Video Studio

The Video Studio makes videos of every kind - product demos, but also brand films, event and booth loops, launch announcements, threat explainers and social teasers - and mixes everything a film is made of: big statements revealed line by line, animated AI shots generated by the video models of your AI providers, extracts of your own footage (an event film, a keynote), demos recorded in the platform's own browser, pictures, slides, title and closing cards - narrated or silent, captioned, branded and rendered to MP4. You make it with your assistant, by hand in the editor, or both at once: the assistant sits beside the stage, and everything it can change, you can change too.

Every video, recording and AI shot you can see is listed under Work > Videos in the sidebar.

Turned on per deployment

The Video Studio is off until an administrator turns it on (ENABLED_FEATURES=VIDEO_STUDIO, see Configuration): until then Work > Videos, the Video Studio tool and the video templates are not offered.

Start a video

New video asks for one thing: what the video is for. It makes any video, for work or just for fun.

What kind of video? (optional). Creative: Just for fun, Short story, How-to, Greeting (a birthday, a thank-you, a farewell). For work: Product demo, Brand film, Event, Launch, Threat explainer, Social teaser. The kind opens the request with its words and shows an example brief (Use this example copies it). A kind for work also suggests the template made for it and the format it starts in (vertical for a social teaser); a creative kind stays in free style.
Describe it: who watches, how long, what it should show - "3 minutes about the company, with extracts of our latest get-together, our events and product demos", or "a 30-second video of a cat astronaut on a candy planet, upbeat music, no narration".
Look: free style until you or a kind ask for a template - no brand, no added cards, nothing locked. Change lists the templates; a look you choose stays yours when you pick a kind afterwards, and the dialog offers the kind's template in one click (Use it).
Format: Landscape 16:9, Mobile 9:16, Square 1:1 or Feed 4:5 - the ones the chosen template makes.
Made by: the assistants with the Video Studio tool, your default agent first when it has the tool, otherwise the General Assistant.

The box at the bottom shows the exact request the assistant receives - the kind of video, then your words. Create with the assistant creates the video (untitled: the assistant names it), opens it in the editor and sends that request as the first message of the assistant's conversation, bound to the new video: it plans the scenes, starts the AI shots, writes the statements and the narration and fills the timeline while you watch. Start from a blank video opens a title card to build on by hand instead. Both create what the dialog shows: a blank video is in free style unless you chose a look, and a video just created opens with a clean quality check, whatever the template.

You can also ask any agent with the Video Studio tool in a normal chat ("Make a 60-second demo of creating a case from an alert in our demo OpenCTI, with a voice-over", "Make a 45-second silent loop for our booth"); its video appears as a card in the conversation and opens in the editor.

Brief it like a director

Name the audience, the one thing the viewer should remember, the length and where it plays (a booth screen, a feed, a meeting). For a demo, name the steps to show: the platform does the camera work - it zooms toward what is clicked and typed, draws the cursor with click ripples, compresses the waits and paces every scene on its narration. For a brand film or an event loop, give the message and the material: the assistant writes it as statements over AI shots, extracts of your footage and the brand's own pictures.

The assistant's know-how comes from the Video Production skill: the arc of each kind of video, how to write a statement and an AI shot prompt, how to use real footage and script a recording of your product, how to build an explainer from your intelligence, to fix what the quality check finds after each save, to let shots and renders finish in the background instead of making you wait, and to generate a failed shot again rather than swap in a still image without telling you. The General Assistant carries it out of the box, and so do, on XTM One, the CTEM Assistant, the XTM-One Assistant and the assistant of each product (OpenCTI Assistant, OpenAEV Assistant, OpenCRQ Assistant, FIMI Analyst), together with the Video Studio tools; attach it to your own agents like any skill (the General Assistant's skills are set under Settings > General Assistant > Skills, see Settings).

The editor

A video opens in its editor, full screen:

At the top, the header, the same bar as a presentation's: the back arrow to your videos, the title (click it to rename the video), Play (watches the current render, full screen - see Watch a video), Render (opens the Renders tab), undo and redo, Studio (shows or hides the assistant), Share, Download and the menu. Until the video is rendered, Play says so.
On the left, the assistant: a chat bound to this video. It knows which video you are editing and which scene the playhead is on, so "tighten the narration of this scene" or "add a callout on the Create button here" needs no explanation. Its saves land on the stage immediately. Open in chat continues the conversation in the main chat; the history menu lists the conversations about this video. Collapse it to give the stage the room; Studio in the header, or the strip left at the edge, brings it back.
In the centre, the preview: the video as it will render, drawn by the same compositor the render uses, from your draft, live. Recordings and AI shots play from their footage, so you see the camera, the cursor, the transitions, the overlays and the captions as they will be. The narration is spoken with the video's voice: a new or changed line is spoken in the background a few seconds after you stop editing (its block on the narration track pulses meanwhile, and the line is paced at reading speed until it is ready), then plays at its real length. The render reuses the voice you heard and does not pay for it again. Deck slides are drawn at render time and show as a card naming the slide. The preview stays steady while the assistant works on the video: what is on screen keeps playing through its saves, a shot generated again keeps showing its current take until the new one can play, a shot seen for the first time shows its poster while its footage loads, and the next scene is loaded before it comes on screen.
The transport under it: play and pause, frame by frame, scene by scene, to the start or the end, the playback speed, loop, mute and full screen.
The timeline: every scene where the render will place it, named by its title or first line, else by the recording or the AI shot it plays (in progress while it is still being made) - transitions show as the overlap they are, a narration longer than its scene shows the scene stretched to it - with lanes for the overlays, the narration and the music. Click a scene to select it, drag the ruler to move the playhead, drag a scene to reorder it, and use the toolbar to cut at the playhead, duplicate, delete, move, zoom and add a scene. A template's intro and outro are drawn too, locked: they belong to the template. Drag its top edge to make it taller (up to a little over half the screen) or shorter; double-click the edge to reset it.
On the right, the inspector:
- Scene: everything about the selected scene - a card's title, subtitle and style; a statement's lines, kicker, figure, smaller lines, chips, reveal and what plays behind it; a recording's or a shot's trim (start and end), speed and held last frame; the narration and when it starts; the transition in; the overlays (lower thirds, callouts, badges, captions, spotlights and pictures - the brand's logo or mark, at a width and an opacity) at their time or on a recording step. A recording also has its post-production: idle moments (speed them up, cut them out, or keep them), the camera (follow the action, still, or keyframes; the deepest zoom; what it zooms on; its pace - Calm, the default, moves smoothly, Dynamic quicker for a short social cut - and how long a zoom takes, to slow every move down or speed it up), the cursor (size, style, click effect), the browser frame (window or full bleed, address bar, margin, corners, shadow) and step highlights. Following the action, the camera opens wide, zooms in on a step, re-frames on the next one (pulling back on the way when it travels far), backs out when a step opens another page or before a scroll, and zooms in again; a highlighted region is framed whole.
- Video: the title, format and frame rate, the template or free style, the theme (colour scheme, backdrop, accents, logo, watermark), the voice-over (speech model, voice, delivery, pace, volume, pronunciations), the captions, the music (a built-in track, an AI track or an upload, and how it sits under the voice) and the default transition.
- Media: add statements, title and closing cards, AI shots (generate one here, it joins the video while it generates), recordings, and images or video files from your computer. In this video lists the recordings and AI shots the video uses; Your library finds any other one by search (the most recent show until you type), to use it here too. A failed shot is not listed: it shows in its scene, with why and Generate again.
- Renders: render the video and follow, play, download or cancel its renders.
Drag the inspector's left edge to widen it, up to a third of the screen, like the assistant's panel on the left.

A video shared with you as a viewer says View now on its card in the library rather than Open in the editor, and plays its current render across the page, as its public link does: the video with its voice, music and captions, its length, its size and when it was rendered. Until it is rendered, the preview, its transport and the timeline take the whole screen instead, under a line saying it is a preview of the edit. The header keeps the way back, the title, Play and Download once the video is rendered, and the menu (copy the link, duplicate the video). There is no assistant, no inspector, no undo and no editing tool on the timeline. Duplicate the video to edit a copy of your own.

A setting left empty uses the template's value (or the default); a setting the template locks is shown with a lock and cannot change. The title and closing cards a template adds are drawn dashed on the timeline: select one and Customize the opening card (Customize the closing card) gives the video its own copy, with the template's words, that you edit like any scene - its words, its length, its style, the logo - while it keeps the template's look for what you leave empty. To change only how long a card the template adds lasts - what the quality check asks when its words are on screen too briefly - or the small line above its title (its kicker: other words, or none at all), the assistant sets it on the video (template_scenes) and the card stays the template's. A card shows its brand once, in its logo: a kicker the template adds is left out when it would only repeat the name the logo already spells. Every edit saves itself a moment after you stop; Ctrl+Z undoes it (an edit the assistant saved too), and a problem the video cannot be saved with is shown under the header without losing your draft. After each save, the quality check line under the header lists what the audience would see go wrong, if anything (see Good to watch by construction); click a finding to open its scene. When an update of the platform changes what the check finds, an editor left open reads it again rather than keep showing the findings and advice of the older check. Fix with my assistant, on that line, opens the assistant beside the stage with the findings written in its message box, the way the assistant reads them after its own saves: read them, add to them if you want, and send. This is the way to act on a finding you cannot fix by hand, such as an AI shot to generate again.

Key	Does
`Space`	Play / pause
`Left` / `Right`	Previous / next frame
`Shift+Left` / `Shift+Right`	Previous / next scene
`Home` / `End`	To the start / the end
`S`	Cut the scene at the playhead
`Ctrl+D`	Duplicate the selected scene
`Delete`	Delete the selected scene
`Alt+Left` / `Alt+Right`	Move the selected scene earlier / later
`Ctrl+Z` / `Ctrl+Shift+Z`	Undo / redo
`Ctrl+S`	Save now
`L` / `M` / `F`	Loop / mute / full screen

Statements, footage and pictures

A video is more than its recordings:

Statements carry the message in big type, the way booth and launch films do: up to four short lines revealed one after the other (or word by word, or all at once), a kicker above, a figure counted up on screen ("6,000+"), smaller lines under them (a proof point, a translation) and chips ("Open source", "Community led"). Words between asterisks - Too much *noise*. - are drawn in the template's accent colour; the same works in the title of a title or closing card, where a line break you type (Enter in the editor) starts a new line - Uncover threats. over *Take action.*. Two asterisks are an accent only when they hug words: 5* rated, 4* loved is drawn as typed. An asterisk that opens an accent no other closes (We see *everything) is drawn as typed too, and saving the video says which word to close, so the assistant fixes it before a render shows it. A statement sits on the template's backdrop, or over an AI shot, an uploaded video or an image, darkened so the words read (Behind the text, Darken it).
Extracts of your own footage: upload an event film, a keynote or a customer testimonial as a video file (Media > From your computer) and keep its best seconds with Starts at / Ends at.
The brand's pictures: a template carries the brand's pictures, and every built-in video template carries those of its brand's deck - the product logos and marks, diagrams and product screenshots on every platform, and on Filigran Copilot the team and event photos, badges and the partner, integration and customer logos too. The assistant places them as image scenes or statement backdrops, and animates them as figures (a wall of customer logos, an architecture whose boxes wear the product marks). Vector pictures (SVG logos, badges, diagrams) are used like any other.
A logo on a scene: a picture overlay draws the brand's logo or mark (by its role in the template: the logo, its light variant, the mark) or any picture file - an emblem, a product mark - on an AI shot, a recording, a statement or a card, at a corner or wherever it fits, at a width and an opacity, for the time you set. It sits above the scene's picture and below the captions, and is placed like a badge: away from the scene's text and the subtitles. Ask the assistant ("put our mark in the top right corner of every AI shot") or add a Picture overlay in the inspector. A video model is never asked to draw a logo: it would misspell it. A template can add its mark to every scene of a kind by itself (its rules, scene_defaults); the built-in templates do not, their watermark already carries the mark.
A real picture inside a picture: to show your real product on the screen drawn in a still (a booth screen, a monitor, a phone), or your real wordmark on a backwall, the assistant draws the screenshot or the logo into the picture - flat, or in perspective on the screen's four corners - and shows you the result before using it.
Your real screen inside a moving shot: when an AI shot shows a monitor, a billboard or a poster, a screen insert puts your recording, another shot, a video file or a picture onto it, in perspective and frame by frame, so its words stay sharp at the quality you render at. It works on an AI shot or a video file, a recording, or the backdrop of a statement, and can glow a little like a lit screen. Ask the assistant ("play my dashboard recording on the monitor of the booth shot"): it reads the screen's corners off the shot. It can also put the template's own logo there (the full logo, or the bare mark: two different files, the one named is the one drawn); a picture with transparency keeps it, Fit inside the frame keeps the picture whole with the surface showing around it (never a black card) - on a chroma-keyed screen the bands it leaves beside a screenshot or a recording are filled with the picture's own edge colour or the screen's black instead, so the green never shows there, while a logo on transparency keeps the board around it - and the glow halos the picture's own shape. What passes in front of the surface stays in front: ask the video model for a flat green screen and set the insert's Chroma key to Green screen (its tolerance and feather beside it), so the insert is drawn only where the screen shows green and a cat or a hand crossing it stays in front, its green fringe removed - hair included, whatever the feather, while a colour of what stands in front itself (a green-lit face, a green collar, the pink nose of an animal crossing the screen, however little of it a frame shows and however blurred by its motion) is kept and never counted as a fringe: those colours are read on the whole subject, over the whole shot, when you save. The screen's green that fur, hair or a sleeve carries on its own edge in the shot is taken off that edge too, even when the model painted its green leaning to cyan, without touching the colours inside it, and the edge then picks up a little of the inserted picture's light, as a real screen gives it; the insert's despill sets this (edge, on by default; width_px, how far past the edge it reaches, 2 by default; strength, 0-1; own_hues, colours of what stands in front to keep as its own, by name - red, orange, yellow, green, cyan, blue, purple, magenta - or as #rrggbb, e.g. ["magenta", "#e8a0b0"]), edge: false leaving that edge as the shot shows it. When you save, the assistant is told how much of the key the edge of whoever stands in front holds on the frame the insert comes in on, and how much of it the despill leaves, before any preview - and, when what stands in front holds the screen's colour or its opposite (a pink nose in front of a green screen) over the shot, that this is its own colour, kept and never counted. Green screen (like blue and magenta) means the green the shot's screen actually shows - read off the frame the insert comes in on, since video models paint their own shade of it - so the inserted colours come out as they are; a custom #rrggbb key is used exactly as given; without a key, what enters the surface through its edge stays in front of it, while what the model drew on the screen itself is covered. A keyed screen never blinks: on a still shot (Follow the camera off) the screen is filled once from its first frame rather than keyed again on every frame - what crosses it still stays in front, its edge softened by the key's feather, and the tolerance decides how much of the shot's green around the bezel the key takes -, and on a moving one the key is read over five frames together, so the edge of the insert does not jitter with the footage's compression. The glow is a soft halo in the picture's own colour, kept to the screen's bezel: it never lights the wall around the monitor, and a keyed screen glows only when you ask it to. Room light (15 % by default) is how much of the shot's own light and reflections play over the insert - a window's glare across the monitor, a lamp's falloff - laid on as light, never as a tint of the room's colour or of the green screen: a screen lit evenly adds nothing, and 0 draws the picture exactly as it is. The editor's preview plays the insert as the render draws it - keyed, following the camera, the recording moving - once it has been drawn in the background (the scene says it is being prepared meanwhile, and shows the inserted media mapped onto the surface); the scene's thumbnail in the timeline shows its first frame (never the green screen it covers); the inspector lists the insert with its fit, glow, room light, key and timing. On footage whose camera moves, turn on Follow the camera (or ask the assistant): the screen's corners, read off the frame the insert comes in on, are followed in perspective through the footage - by the surface and the frame around it - so the insert stays on its surface. When the surface is lost (the camera cuts away, something covers it), the render and the scene preview fail with surface lost at t=... and the frame where it was lost, rather than draw the insert off its surface. On an AI shot whose camera was locked off after it was made, the insert is followed through the shot as the model made it and placed where the locked-off shot shows the screen, so it does not jitter; when locking the camera cropped into the screen, the render fails as "screen cropped" at that frame instead of drawing half an insert (generate the shot again without stabilisation, and the insert follows its camera). A wall of screens is one insert: ask for several panes on the one surface ("show OpenCTI, OpenAEV and XTM One on the three screens of the wall"), side by side or in rows and columns, each showing its own picture or video with a thin gap between them; the panes follow the wall together, in its perspective, and share its key. The corners you give are a starting point: on every frame the render reads the edge of the screen's bezel around them and snaps the insert to it, so a quad a few pixels off still sits square on the monitor and follows it as the camera moves, and the insert is drawn sharp at the quality you render at (never drawn on the shot and enlarged with it). On a still shot you can leave the corners out altogether: when the video is saved, the screen's bezel is found on the frame the insert comes in on - the flat green, blue or magenta screen the shot shows, in the key's colour when the insert has one - and stored as the insert's corners, so the first preview already draws it on its bezel; the save is refused, that frame attached, only when it shows no such screen. Corners you give within 2 % of the frame's width of the bezel are snapped to it when the video is saved; further off, the save says so and you leave them out to take the bezel. One insert per scene can find its corners so, and an insert that follows the camera takes corners given. A curved screen - one whose edges bow by more than 5 % of their length, the points along the edge confirming it - cannot take a flat insert (each edge is judged only on the part nothing stands in front of: an edge mostly hidden by a person or an object is said as covered and not judged, so a subject standing at the bezel never makes a straight screen look curved): a shot meant to carry an insert that starts from a picture with such a screen is refused before anything is paid, a take that bends its screen fails, and the save says so of an insert on it - generate the shot again with a flat screen. A render that finds a screen curved where the shot's own check passed it on the same take says it disagrees with that check, as a warning that never holds the render. A flat screen seen from the side or from below is fine: the insert is drawn in its perspective exactly, and the assistant only says how far it is turned ("seen at 35 degrees"). Only a shot that carries an insert is checked - a screen replacement, a shot the assistant says will carry one, or a shot a scene already draws an insert on - so a mug or a lamp in an ordinary shot never stops it, and the assistant can say a shot carries no insert at all (then no screen is checked on it, and its stray screens are not cleaned). A rounded or glowing object of a screen's colour is never taken for one: a screen has four straight edges around a flat key colour or a dark fill, and is never the brightest thing in the picture. Under a chroma key the screen is painted flat behind the insert - nothing of the green shows through -, the matte is steadied from frame to frame so it never flickers, the key's shade and how unevenly the screen is lit are both read off the shot, and the green fringe on whoever stands in front is taken out without turning it magenta. A screen partly out of frame keeps its true corners outside it and the insert shows through the visible part: never squeeze the quad to the frame's edge, which skews the insert (the save says so as advice). The save also says when an insert's text would be too small to read: a recording, an uploaded screenshot or an uploaded video whose smallest text is drawn under 10 pixels of a 1080p frame is said as advice, under 6 pixels as a finding the render waits for - typically a wall of many panes, each showing a full-size recording. Both offer only what brings its text to 12 pixels and keeps the page's own layout: the share of the frame the insert needs (fewer panes on a wall), or a camera push on the insert - ask the assistant to "push in on the alerts panel": the insert's zoom draws its picture up to 4 times larger within its surface, easing in between two moments and panned to the region you name. A larger surface or a push is offered only at a size the picture stays sharp at in the quality you render the video at (its latest render's, High before the first), so a push the save suggests is never one the render refuses. When the picture cannot be drawn that large, the advice names what can: the pixel ratio to record it at again (the same layout, more pixels), the picture size to use, or, for a recording, the largest push it stays sharp at and the browser zoom to record it at again so its text is larger in the same pixels. A push sharp at 1080p that a 4K render would refuse says so. A picture drawn under 0.3 of its own size shows as a whole page - a dashboard on a wall screen -, which a push would crop: it is offered no push, and the advice says why. Recording it again with its text larger for its width - on a screen under 40 % of the frame's width the browser zoom (the page zoomed in, its window and frame keeping their size), on a wider one a narrower window - is suggested only when the recording's page keeps its layout at the width it would then be laid out at: once a recording is made, its page is narrowed step by step to find where its layout changes (a menu collapses, cards stack, the page scrolls sideways), and the recording says how far down it holds. A dashboard told to record again zoomed in, or in a window, too narrow for its desktop layout would come back as its mobile layout, cropped and over-zoomed on a wall screen. How much the source is scaled down is given with it, but is not the test: a recording at pixel ratio 2 is drawn at under half its size and stays readable when its text is large enough. AI shots and logos are not judged, nor an insert that is scenery: three monitors sharing one frame show their recordings far too small to read by design, so the assistant marks each such insert ambient and the check says it was not checked, rather than advise on it. Marking an insert so changes nothing it draws, and keeps the scene's preview. The advice always offers to mark an insert ambient when it is scenery. The assistant is told the advice once per insert - the panes of a wall together -, not on every save; the finding, every time. The render and the scene preview also fail an insert drawn over what stands in front of its surface (insert covers foreground), a picture kept whole with Fit inside the frame that has a letterbox baked into it (insert padded: bands on two opposite sides in a colour that is not the picture's own sides - a transparent margin, or a plate the logo sits on, is never one, and Fill the frame is never judged), a logo inserted that stops showing (logo lost), an insert whose colours came out different from the picture itself, keyed or not (insert colour shift: its blacks lifted by more than 5 % of the range, or its colours tinted or turned by more than a few degrees - "tinted blue, with its blacks lifted 18%" - with its luminance, saturation and contrast against the picture, on five frames or more; lower its Room light when the room's light lifts its blacks, otherwise the finding names what shows through it; read away from whoever stands in front and its soft edge, and not read at all on a frame they cover more than 30 % of the screen, the answer saying so), an insert that could not be drawn at all - a logo the template has no file for, an SVG that could not be drawn - which would leave the video model's own drawing on screen (insert missing, with the reason), a Fill the frame fit that cuts the recording's own content - a sidebar, a right column - to fill the screen (insert cropped: use Fit inside the frame, or a screen of the recording's proportions), a quad more than half a percent of the frame's width off the bezel the shot shows - on average or on any one side - on five frames in a row (insert off its bezel, with what fixes it - leave the corners out so the save measures the bezel, follow the camera when the screen moves, or generate the shot again when the screen is curved: on a still shot, a frame or two where the render caught a line drawn on the screen for its bezel keep the bezel of the frames around them, and fail nothing), the key's colour left inside or just around the surface (residual key: the quad stops short of the screen, or the key does not take the whole screen - the finding says where it sits, inside the screen, on the screen running past the quad, or past the bezel, and on which side; when the shot's green panel runs past the monitor's bezel, the key takes that green too at its tolerance and paints it in the bezel's colour, so the shot's own green never fails a take; the key's colour is looked for inside the screen and half a percent of the frame's width around it, and fails only past 0.3 % of the screen on five frames or more - under it, a few pixels nobody sees, the answer says it and nothing fails -, while the shot's own green farther out, its screen's glow on the bezel or a spill on the desk, is taken out of the picture on every frame instead, the answer saying how much; green the inserted picture itself shows - a chart, a ring, a button of the product's own colour - is its own and never counted, the answer saying how much of the screen it covers, and the half percent around the screen shows the bezel's own colour), a matte or a despill that flickers from frame to frame (insert flickers), an insert drawn soft (insert soft: a source with too few pixels for the screen it is drawn on at the quality you render at - a 1280-pixel recording stretched over an 1800-pixel screen of a 4K frame reads soft; the finding names the recording and the size it needs, for a recording the pixel ratio to record it at again - or an insert region that comes out blurred against its source) and a green or magenta halo along whoever stands in front on five frames or more (insert halo, measured after the edge despill and said as the edge spill it leaves at its width, with that edge cut at 100 % and the advice to widen the despill - or turn it on - before generating the shot again; where that edge crosses green the picture itself shows, it is the picture's; the colours of what stands in front itself are left out and said, and a magenta the compositing itself made on that edge is said apart, so the edge's length, the spill, the subject's own colour and the compositing's share read apart without opening the crop; when two previews at different settings leave the same halo, the finding says it is no setting but the subject's own colour, named, to declare in the insert's despill.own_hues - or to accept), each with its time, its scene, the frame that shows it, its value against its threshold, what was left out of the measure (the picture's own key colour, whoever stands in front, the bands beside a picture kept whole) and one fix - never a key tolerance or feather to try, which does not move these measures. Every check reads only the pixels it owns: whoever stands in front of the screen - from its first frame on a still shot too, where they stay in front of the insert - is left out of the bezel's edges, the colours and the key's colour alike. The save, a scene preview and a render measure an insert the same way, at any resolution: when a render finds what the save or a preview of the same scene, unchanged, passed, its finding says it disagrees with that save or preview and is a warning, never what holds the render. The answer to a preview or a render also gives these numbers for every insert - how far the quad was from the bezel, the residual key, the flicker, the sharpness, the halo, and its colours at its room light (luminance, black level, tint or hue, saturation, contrast), so the light can be tuned - and attaches the insert's region cut at 100 % of the rendered frame at its first, middle and last frame - the first and last frames its scene is on screen alone, after its transition in and before its transition out, never a frame blended with the scene before or after, and the assistant says where they were cut - and, for a residual key, the side of the screen where most of it sits -, so you judge the compositing on the real pixels rather than on a downsized frame; the frame of each insert's worst finding is attached too, a failed render's included - at most twelve pictures an answer, the others named so the assistant opens any of them when it needs it - and the assistant opens all of them again like the scene's own frames, and can cut the same region at any other moment of a render, at 100 %. It says which resolution the numbers and the crops come from: a preview's are its draft's, and a 4K render measures and crops them again on its own 2160p frames. Recordings are made at a pixel ratio of 2 by default, sharp enough for a 4K render on most screens. The check for a soft insert runs when the video is saved and with every scene preview, for each quality you may render at (standard, high and ultra): a recording too small for its screen in a 4K render is said then ("soft at ultra"), with the size it needs, as advice that refuses nothing - a scene preview is drawn at 1080p and would never show it. Ask the assistant to record it again for that screen at that quality: it reads the screen's size from the saved video and picks the window and pixel ratio, widening the window and zooming the page in as much when it must, so the page keeps its layout. A whole render at a quality where an insert would be soft is still refused before it is paid for. A whole render also keeps the first, middle and last frame of every scene with a screen insert or a replaced screen at the render's own size - a 720p preview hides what a 4K frame shows - and the assistant looks at them before calling the screens right.
A frame of your footage as a picture: the assistant can keep one frame of a recording (the frame after one of its steps, or at a moment), of an AI shot or of a video file as a picture of its own ("keep the frame of the dashboard step as a picture"), at up to 4K, and use it as an image scene, behind a statement, or as the picture an AI shot starts from. A render of the video is a video file too: the assistant can look at any moment of a finished render before you call it final, whoever rendered it, as long as you can open the video.
Slides: a deck slide plays as a picture with a camera motion; chrome: false (ask the assistant) draws it without the deck's footer, logo and slide number, for a slide shown inside the film rather than presented. When subtitles are shown over a slide, it is scaled to stand above them, so they never cover its footer.
Animated figures: one diagram of a deck - an architecture, a process flow, a timeline, a roadmap's phases, a chart, a diagram, an org chart - drawn alone in the video's colours and built in step by step: an architecture's zones first, then each node on its own, row by row, each link drawn once both its ends are there; a process flow's steps and a roadmap's phases in order; chart series growing; an org chart level by level - the root, then each level's cards, then its teams' bubbles growing to their size, every headcount counting up, the dotted lines last. Each step fades in as it rises into place, and the camera moves to it before ending on the whole figure for the scene's last two seconds. Built, the figure keeps moving to the end of its scene: it drifts in slowly, a light pulse runs along each link from its source to its target, and the core node's glow breathes. Ask the assistant for it ("show the architecture of slide 4 building up while I explain it"); it takes the block from a deck, or writes one for the film. In the editor, the Animated figure section sets how long it lasts, whether it builds step by step or all at once, whether the steps come in evenly over the scene or as the narration says their names (measured on the spoken voice once it is generated; a name is recognised however the narration writes it - in any case, with or without a hyphen or a possessive, in the plural, in one word or two, or by its product name alone, as "OpenCTI" for "OpenCTI - Threat intelligence"; a subtitle shows its whole line at once, so a step named in it comes in when that line appears, never after it, even when the voice reaches the name later in the sentence; the steps it does not name are spread evenly around those it does, up to the end of the scene), how long each step stays on screen before the next one (Time on each step, 0.8 s by default: steps a short scene has no room for come in a few at a time), whether the camera follows the steps or keeps the whole figure still, and what the figure does once built (Once built: keeps moving, the default, or holds still). A figure built with its narration stays on screen long enough after it for the step named last to come in when it is said. The figure keeps clear of the subtitles' band. When the assistant saves a video, it is told when each step of each figure comes in and which words of the narration bring it in. The editor shows the figure built in whole on the scene's thumbnail and when you select the scene; it is drawn in the background once the video is saved - saving never waits for it - and the preview says Figure being drawn until it is; a figure nobody opened for a while is drawn again the next time the video is opened. The figure's build then plays on the stage where the scene plays: move the playhead through the scene, or hover the scene in the timeline to watch it build under the pointer.
Pictures brought to life: turn on Animate with the video model on an image scene (a product screenshot, a team photo, a diagram), and describe how it moves if you like ("a slow push in, the charts light up one after the other"). The video model makes a silent shot from that very image when the video is saved; the scene plays the shot once it is made, and the still with its camera motion until then - or for good when no video model is available, which the assistant tells you. A picture no video model can read is refused when the video is saved, never discovered at render. The assistant animates the stills of a film by itself when AI shots are available. A photo with words in it - a team holding a banner, a sign, a caption - keeps them: ask the assistant to keep that region as it is, and the photo's own pixels are drawn back over it in every frame of the shot, following its camera, so the words are never redrawn. A shot whose camera cannot be followed fails as kept region lost and can be generated again with a calmer move; keep a region nothing in the shot moves across.
Screenshots in motion, without a model: a video model redraws the words of a product screenshot, so a screenshot is animated with effects drawn on the real picture instead, its text untouched: panels of the interface lifted off it in depth as the camera moves (parallax), a band of light sweeping across it, a glow pulsing on one region - a gauge, a score, an alert. Ask the assistant ("lift the alerts panel and pulse the risk score"); the preview plays the effects exactly as the render does. A picture moves by the video model or by effects, not both.

Motion design

Everything a video draws - its cards, statements, overlays and motion graphics - moves the brand's way, set once for the whole video under Video > Motion:

Personality: Precise (crisp, measured moves - the platform's own motion), Calm (slow, soft reveals), Energetic (snappy springs, titles letter by letter), Cinematic (long blurred reveals and a light sweeping across the accent words) or Playful (bouncy springs). It sets every timing and easing at once: how a title comes in, how far a line rises, how fast the lines of a statement follow each other, how a card's logo appears. Speed makes the whole video's motion faster or slower, and Titles come in by overrides how titles split (lines, words or letters). A template presets its personality, and a brand template locks it.
Motion graphics: a Motion graphics scene (Add scene in the timeline, or Media > Cards) lays out a whole animated scene from a few words - a kinetic title, the brand's logo revealed, a figure counted up ("6,000+"), a list built item by item, a statement split across windows, a quote, a title over a moving constellation - in the video's colours and personality, whatever its format: a 9:16 version of the same scene stays readable. Choose its Composition in the inspector and write its words. The assistant can also draw a motion scene of its own: text set letter by letter, shapes and lines drawn stroke by stroke, one shape turning into another, pictures and logos moving along a path, a 3D camera moving through the elements, particles, and Lottie animations exported from After Effects (Bodymovin JSON with its images embedded and its text as glyphs; a file that loads anything from outside itself is refused when the video is saved, and expressions are not played - the animation plays its keyframes).
On the beat: with music, Cuts on the beat makes each scene play a little longer - never shorter, a second at most - so the cut lands on a beat; Cuts and builds on the beat also brings a statement's lines and a motion scene's elements in on the beats. A built-in score's beats are known from its tempo; an AI track or an uploaded sound is listened to once, and a sound with no steady beat is left as it is, the render saying so.
Motion blur: what moves blurs as it would on film. It is drawn at the High and Ultra qualities only, so previews stay fast, and a render with it takes about three times as long.
Figures in space: an animated figure can stand in 3D while it builds (In space: leaning back and turning towards the camera, or turning slowly side to side).

Every frame is still drawn from the time alone, never from the frame before it: a scene preview, a render split across workers and a render drawn again show exactly the same motion.

Good to watch by construction

Every frame is laid out so that nothing covers anything else, whatever the assistant (or you) wrote:

The subtitles keep their band. While someone speaks, the text of a card or a statement, a lower third and a recording's browser window stay above the subtitles (below them when they are at the top) - burned in or not, because a player draws the subtitle track in that same band. A card's whole text, not just its title, is scaled down to fit the room left, and a word too long for the width (a URL) is scaled into it.
Overlays find their place. A badge or a callout asked where the scene's text already is takes the first free corner; two callouts on one element stand on either side of it, or one above the other; a lower third stands above the subtitles.
Subtitles are readable. A subtitle holds on two lines, set a little smaller when its sentence is long.
The watermark leaves the stage to the brand. Title and closing cards already carry the brand: the corner watermark leaves while they play and comes back with the content, a moment after it starts and once the card has left the frame, however long the transition. A statement does not draw its logo while the video has a watermark: the watermark already shows the brand, and the logo shows again once the watermark is off. The brand mark is never drawn twice on one frame. The brand film templates open on a centred title card with the logo.
Transitions cut like an editor. Left on Automatic (the default, and what every built-in template uses), each cut is picked from the scenes on either side: the opening card pushes into the video, words dissolve into pictures and pictures into words, a run of AI shots dissolves from one to the next (or, at a Dynamic rhythm, alternates a push through the frame and a whip pan), the closing card comes after a dip to black, and a shot too short to dissolve is cut. Video > Transitions sets the rhythm - Calm, Dynamic, or Classic, which only cuts, crossfades and fades to black, the grammar of a brand film (the brand film templates use it) - or one transition for every cut - cut, crossfade, fade through black, slide, zoom, blur, wipe, whip or flash - and a scene's Transition in overrides it for that scene, which says what it comes in with. The assistant is told the transition planned into every scene each time it saves or reads the video, so it can check the cuts before rendering.
Footage keeps moving to the end of its scene. A scene lasts as long as its narration. When the narration is longer than the AI shot, the animated picture or the video behind a statement, the footage plays slower to fill the scene, down to 70% of its speed. A shot whose own sound you keep plays at its speed, because slowing it would slow its sound. A recording first plays the pauses it sped up more slowly, never slower than twice real time, so a page loading still reads sped up; the steps keep their real speed, and pauses set to Cut stay cut. Whatever is still missing holds the last frame, drifting slowly in until the next scene, so the picture never stops dead before a transition. Your own Hold the last frame is kept as you set it: a still moment, which does not drift. The preview and the render play it the same way.

What a layout cannot fix is measured: after each save, the video is played in the render's compositor at the moments that matter and checked for text covering text, text cut by or touching the edge of the frame, text scaled down so far it gets hard to read, text still appearing when its scene ends, words on screen too briefly to be read, subtitles too fast to follow, words over a backdrop not darkened enough, overlays too short to read, and the brand mark drawn twice on one frame (brand mark drawn twice: the watermark, a card's logo or a picture overlay of the brand's mark together, with the time of the frame and what to switch off to keep one). The assistant reads the findings with the result of every save and fixes them before it renders; you see them in the editor under the header.

The pictures and the sound are checked too, scene by scene, each finding saying what to change: an AI shot that failed (the render would refuse it) or is still being made, a ready shot being generated again or whose last regeneration failed (advice only, said once whatever the number of scenes showing it, naming the take that plays: its previous take plays, in a render too), an animated image whose animation failed or is still being made (advice only: the scene plays the still image), a shot asked with a still camera whose camera moved (camera moved on a locked-off shot; a real product screen drawn into a shot follows the camera, so it is not flagged for it; set motion: intended on the clip scene when the move is wanted, and a preview of every scene showing the shot that passes clears it, the check naming that preview and its revision), an image smaller than the frame it is drawn on (it would look soft), a screen insert whose corner leaves the frame (advice: the screen keeps its true corners and the insert shows through the visible part - a quad mostly outside the footage the frame shows is a failure; the render and the scene preview also check every frame an insert draws, above), music that ends before the video with Loop off, music too quiet to be heard (silent, a long quiet passage, or lost under the voice - advice only, see "Music"), a narration no voice will speak (the voice-over off, or no speech model available), a narration spoken by another voice than the video's because the video's voice failed it (see below), a card or a statement that writes the brand name its logo already spells (Filigran in the kicker under the Filigran logo: the audience reads it twice), and an AI shot or a recording still holding its last frame for more than a second and a half under a longer narration (advice: generate a longer shot, continue it with a shot made from its last frame, record more of the flow, or shorten the narration). A shot or a recording shorter than its narration says the speed that makes it last as long, beside Zoom to fill the scene, and a screen insert running past the point where a looping shot starts again says where that seam falls and the length that ends the insert before it. When a recording is made, the assistant is told how long it plays at each speed and how much narration that carries, before it writes one. These are read from what the platform knows of your media, so they are shown even where the layout cannot be measured, and the editor reads them again when a shot is generated again, without waiting for the next save.

Before a scene preview or a render starts, everything it would find without drawing a frame - a screen insert off its bezel or on a screen that is not flat, footage shorter than its narration, a frozen last frame, an insert past a loop's seam, text that does not fit or covers other text, a shot or a recording not ready - is checked at once, and the start is refused with the whole list in one answer, each finding saying first when only a new shot fixes it. Corners a little off the bezel are snapped and saved as a new version of the video before the start, said in the answer.

Templates

A video template keeps every video of a team to the same look. It is a template like the ones that dress presentations - listed under Agent Studio > Templates, shared, duplicated and published the same way - and it carries:

what it is made for: product demo, brand film, event, launch, explainer or social - New video suggests a template for each of these kinds of video;
the palette, fonts, logos and pictures of the brand (the same colours as the brand's deck);
the look it presets: the backdrop, the watermark, how its videos move (see Motion design), the transitions, the captions, the voice-over style, how a statement reads, how a recording sits on screen (the browser window, the cursor, the camera). A preset fills only what the video leaves unset: the editor's Watermark field shows the template's corner until you pick another one or None, and when a video draws no watermark while its template draws one, saving it says so;
the art direction of its AI shots: the look every AI shot made for the video takes (the light, the lens, the materials, the palette) and what no shot ever shows - see AI shots wear the template's art direction;
the cards it adds: a title card before the first scene and a closing card after the last, when the video has none of its own - a starting point, never locked: a video's own first title card (last closing card), written in the editor or by the assistant, replaces the template's and takes its look;
the choices it locks: a locked choice always takes the template's value, whatever the video asks for; the editor shows it locked, and the assistant is told so when the video is saved.

These templates ship with the platform:

Template	Where	Made for	What it does
Filigran Product Video	Filigran Copilot	Product demo	The Filigran palette (dark or light) on the aurora backdrop, the Filigran mark as a watermark, every recording in the same browser window, a Filigran title card and a "Book a demo" closing card. Locks the colours, the backdrop, the watermark, the captions and the browser window.
Filigran Brand Film	Filigran Copilot	Brand film, launch	Statements revealed line by line over AI shots and event footage, the brand palette on the aurora backdrop, the Filigran deck's pictures and logos, cuts, crossfades and fades to black only, and the "Uncover threats. Take action." closing card.
Filigran Event Loop	Filigran Copilot	Event	A silent loop for a booth or stage screen (16:9 or vertical for a totem): no voice-over, no captions, short statements alternating with AI shots and product screenshots, the tagline at the end.
THREAD Conference Video	Filigran Copilot	Event, launch	The THREAD Color Set, dark like the stage, every card and recording drawn over the THREAD pixel mosaic, Elms Sans titles, the mosaic emblem as a watermark, the edition on the title card and a "See you there" closing card under the THREAD logo. Formats 16:9, 9:16 and 1:1.
XTM One Product Video	Every platform	Product demo	The XTM One palette on the mesh backdrop, recordings in the same browser window, an XTM One title card. Locks only the colours and the browser window.
XTM One Brand Film	Every platform	Brand film, launch, event	Statements over AI shots and real footage, figures counted up, the XTM One palette on the mesh backdrop, cuts, crossfades and fades to black only.
Threat Briefing	Every platform	Explainer	One threat, campaign or vulnerability explained from your intelligence: figures and ATT&CK techniques in statements, your intelligence platform's own screens as the evidence, AI shots that set the scene, a calm analyst voice-over with captions, a "Threat briefing" title card.
Social Teaser	Every platform	Social, launch	15-30 s, vertical first: a hook in the first two seconds, words revealed one by one, quick cuts, captions burned in for muted viewing.

Every one of them directs its AI shots: filmed rather than rendered - natural light, real places and materials, the brand's own tones, no lettering, logos or invented screens - and never laser beams, starfields, lens flares, glowing networks, holograms or the other stock images of a generated "cyber" picture. They never limit how many AI shots a video has: a video made of AI shots, product demos and animated pictures alone is a complete video.

Each one also moves its own way (see Motion design): the product videos are Precise, the Filigran one revealing its logo from the left and sweeping a light across its accent words; the brand films are Cinematic, cut on the music's beat, with motion blur; THREAD and the event loop are Energetic, on the beat, with motion blur; the threat briefing is Calm; the social teaser is Energetic and a little faster, its titles word by word. The Filigran templates and THREAD lock their personality; a video still sets its own speed, beat sync and motion blur.

A video an assistant creates in a chat without naming a template wears the platform's default one: the Filigran one in Copilot mode, the XTM One one elsewhere. Free style (the look New video starts in, also under Video > Template in the editor) applies no template at all: no added cards, nothing locked, every choice is the video's. A video keeps its look when the assistant rewrites it: its template, or free style, until a template is chosen.

To make your own, duplicate a template and change it: its palette in the Design editor, its rules under Configuration > Advanced (JSON) in extra_metadata.video (see Templates). A template change restyles the next render of every video wearing it.

Recordings and browser profiles

A recording is captured in the platform's own headless browser - nothing to install - at a high resolution (2400x1350 by default), with human pacing: the cursor glides to each element, typing has a natural rhythm, scrolling is smooth. Each recording keeps its action log (what was clicked, where, when), which is what the camera, the cursor and the callouts are drawn from. Sensitive parts of a page can be blurred in every frame or hidden (a cookie banner). Recordings are made by the assistant from a step script: Record a new flow with the assistant in the editor's Media tab asks it. A step marked optional (a banner that may not show) gives up after 3 seconds instead of 15, and the time it spent looking is cut from the video, whatever the idle setting. A step choosing an option works on any list, not only a classic drop-down: a picker such as a tenant or organisation selector is opened, the option is found by the name it shows (typed first in a picker you search in), and the cursor moves to it and clicks it; the camera frames the picker and its choice. An option the list does not have stops the recording naming the options it offered. The contact sheet the assistant checks a recording with - and every poster or screen it shows you - is a file of the conversation, so it opens again when you come back to the conversation later.

To record an application you sign in to, use a browser profile (Tools & Integrations > Browser profiles, or Browser profiles at the top of Work > Videos): the start URL, the sign-in URL if it differs, the account to use, and the session its sign-in saved. The same profiles let your agents browse the application signed in; their sign-in, MFA, sessions and timeouts are described on that page.

The assistant looks before it records: it can open a page of the signed-in application in the profile's saved session, without typing anything or running a step, and read its screenshot and the names of what can be clicked or typed in - the names the menu really shows ("Agentic Flow", not "Flows") - so the recording's steps find their elements the first time. A step that still misses its element is reported with the three closest names the page offered at that moment.

Recordings start signed in, or not at all: every recording checks the session before its first step. When it has ended, a profile that can sign in by itself (its password and TOTP secret, or the identity provider's saved session) signs in again first; otherwise the recording stops with profile not authenticated: landed on <page> and makes no video - sign the profile in, then record again. Some applications show their sign-in page for a moment while they read the session, then open: the recording waits for the application to show signed in (the element or the address you set on the profile, else the same check as the session check) and starts only then, so its first frame is the application, never its sign-in page. An application that does not open signed in within the profile's sign-in timeout stops the recording the same way. A recording may use any profile you may use; the profile does not need to be given to the assistant.

Recordings reach public addresses only, unless an administrator allows browser profiles on private networks and the profile declares the hosts.

AI shots, voice-over and music

AI shots are animated footage generated by a video model - an opening shot, b-roll behind a statement, a place or people at work that sets the scene - optionally starting from a picture or from a frame of your own footage, so the shot opens on something real. That frame can be an image, a step of a recording (the frame its contact sheet shows after that step), a moment of an uploaded or rendered video, or the last frame of a recording or of another shot, which the new shot then continues: placed right after that footage, it carries on from where the footage stops. Ask the assistant for it ("start a shot from the dashboard step of my recording"). Before it pays for the shot, the assistant looks at that frame and moves it if it is not the right picture; a page caught still loading (a spinner, a blank page) is noticed, and the shot starts once the page has settled instead. The frame taken from your footage is kept as a file of the conversation: the assistant shows it with the finished shot, the inspector shows it under First frame with where it was taken, and a shot generated again starts from that same picture, even once the recording is gone. A frame of your footage must have the shot's shape: a landscape recording does not start a vertical shot unless the assistant is asked to crop it, which keeps the middle of the frame at the shot's shape (a picture is used as it is). They are scenes of a video like recordings: generate one from the editor's Media tab (or ask the assistant) and it joins the video at once, shown as being generated until it is ready. They use the video model your administrator chose under Settings > AI Models > Video Generation (see Video generation): Veo on Google Gemini, or what an OpenAI-compatible gateway such as LiteLLM serves (Veo on Vertex AI, Runway); the editor lets you pick another listed model for one shot. A model named for a shot, or for a shot generated again, is made by the provider that lists it - never by the one chosen in Settings - and a model two of your providers offer is made once you say which. A model its provider has retired is never offered or used - OpenAI retired its Sora video models on 2026-09-24 - and when no video model is available at all, the assistant says why before it plans the video, and nothing is generated to fail. When a model does not make the size, length or resolution asked, the shot is made at one it does and the shot says so; a provider that is busy or briefly unreachable is asked again, waiting a little longer each time, for up to ten minutes. A provider's rate limit or quota never fails a shot on the spot (see below). A shot is never paid for twice: a generation interrupted by a restart of the platform is picked up where the provider is, and a video already made is kept even if what follows it fails. The platform generates a few shots at a time (four by default, VIDEO_STUDIO_CLIP_CONCURRENCY), so when the assistant asks for more at once, the others wait their turn: the assistant waits for each shot for a few minutes at most (VIDEO_STUDIO_CLIP_INLINE_WAIT_SECONDS), and a shot still being made after that goes on in the background. The assistant is told why - waiting for a free slot, queued at the provider, rate limited - and how far the shot is, and checks on it before rendering.
The same subject in every shot: a shot can be given one to three reference pictures - a character, an animal, a garment, a product, a logo - that the video model keeps the same, wherever it stands in the frame: give every shot of a film the same references and the subject no longer changes from shot to shot. A start frame is for continuity (the shot carries on from a picture), reference pictures for consistency. Only some video models take them (Veo 3.1 and Veo 3.1 Fast, without a start frame and at 8 seconds; Kling 3.0 Omni, beside a start frame; Seedance 2 on Runway, without a start frame); the assistant is told which, and a shot asked with references on a model that takes none is refused before anything is paid. A shot generated again keeps its references. A shot is asked in one of three modes - a start frame only, reference pictures only, or both on the models that take them together - and the assistant is told which each time. Generating a shot again switches between them: giving it references on a model that takes one or the other drops its start frame by itself (Start frame dropped); the assistant can also drop the start frame on purpose (start_frame: false), go back to a start frame alone, which drops the references (start_frame: true with the picture to start from: Reference images dropped), or start the shot from another picture or another moment of its footage. The picture it started from stays in the conversation. Before generating many shots at once, the assistant can check them all first (validate_only): every check runs - the model, the references, the start frame and its shape - and the answer says the mode and whether the provider cools; nothing is generated or paid; a combination the model cannot take is refused at once, naming the model and the two ways out.
Ingredients - your logo, your product, a character: in place of bare reference pictures, a shot can be given one to three ingredients, each with what it is and, if you like, its name ("the Filigran logo"), and it keeps the picture it starts from on any video model. A model that takes reference pictures beside that start receives them; any other model is told to keep each one exactly as in the picture the shot starts from, so that picture must show them - the assistant composes it first when needed. On such a model a logo is also drawn back onto the finished shot with its own pixels: it is found in the picture the shot starts from (or placed where the assistant says), followed through every frame as the camera moves, and drawn over what the model made of it, so it is never misspelt. It is drawn in exactly one place in every frame: on an object that moves past something else (a mug falling over a table), the surface carrying the logo is followed, never what lies behind it. When the object turns its logo away from the camera (a mug that rolls), the logo fades out over three frames instead of being squeezed into a smear, is not drawn while it faces away, and fades back in where it faces the camera again; an object at rest keeps its logo perfectly still. A shot that loses sight of a logo facing the camera (the camera cuts away, a hand covers it) fails as logo lost, its video kept to look at. Every finished shot is checked for its logo showing twice in a frame: such a shot says logo duplicated at t=... with that frame, and a scene preview or a render of a scene playing it fails until the shot is generated again. The shot says which way its ingredients went (sent, named, composited or tracked), and a shot generated again keeps them.
A logo on a subject the shot does not open on. An establishing shot of a building or a street, with the branded subject walking in later, has no logo in the picture it starts from. Give the logo its subject instead: a picture of the subject wearing the logo (a conversation file, or a frame saved from another shot) and, if you like, what to call it ("the cow in the navy blanket"). Once the shot is made, the subject is found in its frames, the surface the logo sits on - the blanket, the garment, the panel - is followed through every frame, and the logo is drawn there with its own pixels, in perspective and under the light of the scene, on every instance of the subject (a herd each wearing it gets one logo each), from the frame it enters to the frame it leaves. This works on every video model, with or without a start picture; a model that takes reference pictures is sent the logo first, and the logo is drawn onto the subject only when that take does not show it as given. A shot whose subject never appears fails as ingredient subject not found; one whose surface cannot be followed past a frame (a hand over it, a fast turn) fails as ingredient surface lost at t=..., that frame kept to look at. The shot says the share of its frames the logo was drawn on. Ask the assistant ("the cow wears our logo on its blanket - it walks into the street shot").
Every shot's frames are kept and measured: once a shot is made, its first, middle and last frames are kept with it - the assistant sees them with the shot, the inspector shows them under The shot's frames - with how close its first frame stayed to the picture it started from, the share of that picture's words it still reads, and how far its last frame moved from its first. Ask for a minimum likeness ("keep the first frame close to the screenshot") and a take that drifted further from its picture fails as drifted from reference, its video kept for you to see, to be generated again in place.
The picture a shot starts from can be any picture: a PNG or a JPEG goes as it is, an SVG (a logo, a badge, a diagram of a template) is drawn first, and any other picture (WebP, GIF...) is converted - every video model receives a picture it reads. One that is no picture at all is refused when the shot is asked for, before anything is paid.
The text of a shot is checked. A video model asked to move a diagram, a badge or a screenshot redraws its labels and sometimes garbles them ("Orcheatration"). Once a shot made from a picture is generated, the words of its first and last frames are read and compared with the picture's - with the words the picture itself reads clearly, that is: lettering that is small, soft or at an angle in the start picture (flags on poles, a distant banner, signage seen edge-on) reads as garbage there and as other garbage on the shot, so it is recorded as unreadable in the start picture and never held against the shot. The shot says which words were checked, which were left out and how surely they read, and - reported, never a failure - any lettering it shows that the picture does not, which is what a drifting model usually adds. When a word is drawn wrong, the picture's own lettering is first drawn back onto every frame of the shot, following the camera, and the shot is read again - once, with nothing new generated or billed. Only a shot that still reads wrong fails as text drift, naming the word, the frame and a crop of it, so it can be generated again (a calmer camera move helps) or the scene plays the picture as a still. Small text the model melts into unreadable marks (a screen in a photo, a badge) fails it the same way. A word the camera moved out of frame is not a drift. An animated image whose animation fails plays as its still image. The reading is done by your chat model; when it cannot run, the shot is kept and says its text was not checked. Ask the assistant to skip the check for a shot meant to change the picture's words.
What a shot shows is checked against its prompt. A shot whose prompt names subjects ("six cows grazing in a meadow") has its first, middle and last frames read once it is made: a subject missing, of another kind, deformed (extra limbs, human hands on an animal), turned into something else, or something the shot was asked not to show, fails the shot as content drift, naming what came out - its video is kept for you to see, and it can be generated again in place. A subject that walks out of the frame for a moment is not a drift, and a shot whose prompt names no subject (a landscape, an abstract motion) is read only for what it was asked not to show and for what its template avoids (said, never failed). A detail nobody notices at full HD - a marker that lost its dot, a seam on a sleeve - is minor drift: the shot is kept and its notes name the detail; ask the assistant for a strict check on a close-up where every detail shows. The reading is done by your chat model; when it cannot run, the shot is kept and says it was not checked. Ask the assistant to skip the check for a shot meant to transform its subject. When two takes in a row show the same thing the shot was asked not to show (a coloured light the model keeps drawing), the next take asks for it in the prompt itself - "No red light anywhere in the shot." - rather than in the negative prompt alone, and turns the model's guidance down where the model has such a setting; the answer says so, and the take is still checked against what the shot was asked not to show.
Screens a shot was asked not to show. A video model asked for an office "without screens or readable text" sometimes draws monitors showing an interface anyway. Instead of failing such a shot, the assistant can ask for those screens to be cleaned: flattened (switched off, keeping the room's soft light on them), or showing a real screenshot you choose. Only a shot whose one problem is what those screens show is cleaned; each screen is followed through every frame, as the camera moves, and drawn over. A screen someone passes in front of is never drawn over (the shot fails as before, saying why), and the cleaned shot is read again: anything else it still shows that it was asked not to fails it.
A real product screen in an AI shot. To film the product on a screen - a laptop on a desk at dusk, a booth wall, a control room - the shot starts from a picture of the room with the screen in it, and the real screenshot is named with the screen's four corners in that picture. When the screen of that picture is one solid green, blue or magenta surface (ask the assistant to draw the room that way), its corners are found in the picture rather than guessed, and the assistant can check them before anything is paid. The video model draws the room, the light and the reflections around a flat screen, and the screenshot is drawn into that screen in every frame, following the camera, under anything that passes in front of it: its text is exact, so it is never checked for text drift. The screenshot is kept whole, letterboxed in its own background colour where the screen has another shape (a sidebar or a right column is never cut; asking to fill the screen instead is refused when it would cut the content), and it is carried to the screen's bezel, so no line of green stays between it and the frame around it. A shot whose model did not keep the screen flat (a fast camera, the screen leaving the frame) fails as screen lost, its video kept to look at, and is generated again with a calmer move; so does one that leaves green around the screen (key spill) or whose screen edge jitters from frame to frame (key flicker). A camera move the model adds anyway is not a failure while the screen is followed. A camera move is measured on the shot's frames, never guessed from a still picture: to hold a shot to a still camera, ask the assistant for a camera tolerance, and a take whose camera moves more fails as camera moved, its video kept. Some video models do not hold a locked-off camera whatever the prompt says (Veo from a prompt alone drifts, from a start frame it holds; Kling holds it): with a tight tolerance (0.5 or less), a new shot is made with a model that holds it, and when you named the model, the assistant says before anything is paid that it will not hold the camera, and what would. Ask the assistant ("show our dashboard on a laptop in a dim office").
A drifting camera is locked off. A video model asked for a still camera often lets it drift a little - a slow push, a few pixels of pan. When a shot is asked with a camera tolerance, or its prompt asks for a still camera, the finished take is stabilised before it is measured: every frame is locked to the first one and the edges the drift uncovers are cropped away, so a take that only drifted is kept instead of failed. The stabilised take is kept only when it really moves less; otherwise - the camera could not be followed, the drift is too large to crop (more than a fifth of the frame), or locking it would not help - the take stays as the model made it and says why. The un-stabilised take is kept beside the stabilised one, and the inspector says how much the motion went down (Stabilised: motion 0.42 to 0.03). A shot with your real screen drawn into it is never cropped into: when no crop keeps that screen whole, the shot stays as the model made it and says so. The assistant can also ask for it on any shot, or turn it off for one.
Another model when a take fails. The assistant can give a shot one to three other video models to fall back to. When a take fails a check of its frames - its camera moved, it drifted from its start picture, its text or its content drifted, its screen or its logo was lost - the shot is generated again in place with the next model, without waiting to be asked; each model is tried once, and each is checked before anything is paid (one that cannot take the shot's reference pictures is refused, and so is one its provider no longer serves - it answered "not found" for it - named in the answer). A model a provider lists but no longer serves is left out of the models the assistant offers, with the reason, from the moment the provider says so. The shot's own model named among them is simply left out, the assistant saying so - its first take is already made with it. The shot then says which model made the take it kept, and why the others failed (Made by the fallback model kling-v3 in the inspector).
Silent shots: some video models make their own sound (8 seconds of synthetic ambience for a Veo shot). The assistant can ask a shot silent - the model is asked for no sound, and any sound it makes anyway is removed before the shot is stored - which is what a shot under a music bed or a narration wants; the shots that animate a picture are always silent.
A provider's rate limit never fails a shot on the spot: a shot the provider refuses because too many were asked at once (its limit per minute) waits and is asked again - after the wait the provider names, else after a short pause - for up to 30 minutes, nothing being made or billed meanwhile, and the assistant says it is rate limited. Past that, it fails saying so. The refusal is the whole batch's, not the shot's: that provider cools until the wait it asked for (or a wait that grows each time it refuses again after a cooling ended - the shots of a batch refused together count as one refusal), and every other shot asked of it meanwhile waits in the queue without being asked - saying until when (Waiting in the queue until ... in the inspector) - and starts by itself once it cools. A quota (the day's, the plan's) is usually counted per model: a refusal for a model's quota cools that model alone, and the provider's other models still take shots. A shot refused for a quota waits up to a day, even when the provider asks for the whole day: a quota is never a failed shot. The assistant is told before it queues a batch (validate_only and the list of video models say which model cools, and until when), so it queues the batch once or picks a model that does not cool.
A shot that fails says why where it plays - the preview, the timeline, the inspector (a prompt the provider's safety system blocked, a model the provider does not serve, a refused key, a text drift). The video keeps saving and previewing around it. Generate again (in the inspector, with the prompt you can reword) makes it again in place: every video using it keeps it. The assistant can also generate again a shot that is ready but came out wrong: the shot keeps playing its take - in the preview, the quality check and a render - until the new take is ready and has passed its checks, then the new take replaces it in every video using it. A new take that fails (refused, drifted, cancelled) leaves the previous one playing and nothing broken: the inspector says The last regeneration failed: the previous take still plays with the reason, the quality check says it as advice, and the failed attempt's frames are kept to look at until a new take replaces it. Nothing already made is generated again to get the previous take back. The assistant can drop the previous take at once instead, when you want it gone whatever happens (discard). A failed shot is never kept as media: it is not listed, and once no video uses it, it is deleted an hour after it failed. Each attempt is a new generation billed by the provider, so one shot is generated again at most five times from the same prompt and start picture; a reworded prompt or another start picture starts the count again, and a retry that only asks another model or a different camera tolerance is never counted. The assistant keeps fixing a shot that failed - a reworded prompt, another model - and stops only when the very same request keeps failing. A video renders once all its shots are ready.
Voice-over is spoken from each scene's narration by the speech model your administrator chose under Settings > AI Models > Speech Generation (see Speech generation): OpenAI, Azure OpenAI, Gemini, ElevenLabs - with your team's cloned brand voice - or what an OpenAI-compatible gateway such as LiteLLM serves. A video, or its template, can name another speech model and voice (Video > Voice-over in the editor), and a delivery style for the models that take one. A video that follows the platform's speech model and has a style is spoken by a model of the same provider that follows the style when the platform's does not (OpenAI's tts-1 and tts-1-hd take no style, gpt-4o-mini-tts does); a video naming a model that ignores its style is told so when it is saved, not after the render. Speech model says which model the platform uses when the video names none; a video that names its own keeps it when the platform's changes, and the assistant says which one will speak each time it saves the video. Each scene is spoken knowing the lines before and after it, so a long narration stays continuous on the models that support it. The editor's preview speaks the narration with the same voice, and the render reuses it; both speak as you, through the speech models you may use. Without a speech model available to you, the narration is shown as captions only, and the render says so.
Pronunciations tell the voice how to say a word it would misread: a product name, an acronym. Your administrator keeps the platform's pronunciation lexicon under Settings > AI Models > Pronunciation lexicon (every save is a new version; an older one can be restored), and a video adds its own under Video > Voice-over > Pronunciations, an entry of the video replacing the platform's for the same written form. An entry has the term as written, how it is said (letters read one by one as capitals joined by hyphens: X-T-M), and optionally its IPA, how it is matched (the whole word in the same case, the whole token only, or any case) and what reads it (the voice, the captions or both: the captions keep the written words otherwise). The Play button beside each term speaks it with the video's voice, so you can choose a voice on how it says your brand before rendering. Each speech model gets the term the strongest way it can take: its phonetic spelling when it reads one (OpenAI's newer voices, such as gpt-4o-mini-tts, take the IPA in their delivery instruction), the provider's own pronunciation dictionary, or the narration rewritten; letters read one by one are always sent as one word - one phonetic token for the letters together (XTM One as XTM said in one breath, then One), or the letters written together - never as letters separated by hyphens or spaces, which a voice reads with a pause after each; when the model can take no IPA (tts-1-hd), the term is said by how the entry says it, and the render's notes say so. Then every narration is heard back: the speech-to-text must write the term as written (case, punctuation and spaces aside: Open CTI is OpenCTI), as the entry says it, or in words that sound like its IPA with at most one sound off. A term the voice says otherwise (CTEM heard as stem, Filigran heard as Phil Igram) fails it even when every other word is right, and so does a term not said as one word - a pause inside it of more than about a sixth of a second (Phil... egran, Open... C-T-I, letters spelled one at a time) fails it as spelled with pauses, whatever speech model said it; a number counts as its words whether it is written in digits or in words, by the speech-to-text or in the entry (XTM 1 is XTM One, 85% is eighty-five percent, an entry said Windows 11 heard as Windows eleven). A narration that fails is spoken again (the same voice with the letters of each term written as one word, then slower, then the closest voice of the same speech model, which the render says) - never placed silently. The finished video's voice is heard once more, after its encoding; the render's checks say, for each term, how often it is said and how often as the lexicon says it, and a term accepted in other words than its entry writes (Filigran heard as Philigran) is a warning with its time, so you can listen to that spot without the render failing. When a video is saved, the quality check warns about a narration that writes a term in another case than its entry (Opencti), unless the entry only changes the case of a word said as written (eXtended said extended), a term the lexicon reads as a word spelled out letter by letter (C-T-E-M), and a brand-like word no entry covers. These warnings are about the narration only: a title or a line on screen is never checked against the lexicon. A template can carry the brand's own list.
The voice is mastered: the silence speech engines leave before and after each line is trimmed, the rumble removed, the sibilants softened, the level evened, and every scene and every speech provider comes out at the same loudness.

AI shots wear the template's art direction

An AI shot is made for a video, and takes the art direction of the template that video wears, whoever asks for it - you in the editor, the assistant, the animation of a picture. The platform adds it to the shot itself, so a prompt only says what the shot shows (the subject, the action, the camera):

The look is added to the prompt of a shot drawn from a description. A shot that starts from a picture or from your footage keeps that picture's look.
What no shot shows is added to the negative prompt of a video model that takes one (it is never written into a prompt: a model draws what its prompt names, so a model without a negative prompt is steered by the look alone). The finished shot's frames are read for it: a shot that shows it anyway is kept - a generated shot is paid for - and says so, and you or the assistant can generate it again in place if it stands out.
A shot made under another template's direction (made for another video, say) is named when the video is saved, as advice: it keeps playing, and generating it again from the inspector (Generate again) or by the assistant makes it under this video's art direction, in place. Shots made before templates had an art direction, free-style shots and shots started from a picture are left as they are.

The art direction never limits how many AI shots a video has, and never stops a shot or a render.

A shot asked for no video at all takes the platform's default template's direction; a free-style video's shots take none. A template's direction is shown on its page and in New video (Directs its AI shots, with what it never shows), and lives in its rules (extra_metadata.video.shots, see Templates): duplicate a built-in template to write your own.

Music

Every video can have a music bed, chosen in Video > Music and sound (or by the assistant) from three sources:

Library: royalty-free tracks the platform composes for each video - nothing to license, no provider needed. A track is written for the video it plays under: its length, its sections changing on the scene changes, its final chord on the last frame; it plays sparser under a narration and lifts on a statement or an AI shot without one. Seven moods: Bright Horizon (uplifting, launches and demos), Deep Focus (calm, narrated walkthroughs), Signal (dark, threat briefings), Ascent (cinematic, brand films), Late Notes (lo-fi, social clips), Momentum (energetic, teasers and event recaps) and Quiet Light (minimal piano, customer stories). Play beside each one plays half a minute of it.
AI music: describe the music (genre, mood, instruments, tempo, how it builds) and Compose for this video: a music model composes it at the video's length, instrumental, with the model your administrator chose under Settings > AI Models > Music Generation (see Music generation). It takes up to a few minutes; the preview plays it once it is ready, and a track that failed says why and can be composed again in place. Your AI tracks can be reused in your other videos. Never name an artist or a song: the providers refuse it. Compose it once the shots are made: a video grows as its shots come out and its narration is spoken, and a track composed early can end before it. When it does, the assistant is told when the video is saved ("Music ends at 142 s of 183 s"), and can compose the same track again at the video's new length, in place; the render's notes say how long the track is against the video. A track is composed 8 seconds longer than the video (and never shorter when it is composed again) - than its current render when that came out longer than planned, the voice taking more time than estimated - so its own ending lands after the closing card and the music fades out on the last frame. Every save says whether the video's current render was made with the music it plays now: after the track is composed again, the render still plays the earlier take until the video is rendered again.
Upload: any sound file of yours.

Whatever the source, the music is levelled against the voice: Volume is its level where no one speaks (1 is as loud as the voice), and it goes down under every sentence (Lower it under the voice, by 10 dB by default) a moment before the first word, and comes back up after the last - never pumping in between. A clip that keeps its own sound lowers it too. An uploaded or AI track shorter than the video repeats with a crossfade; Start at skips its intro, and The track's own end lets its ending land on the video's end. The editor's preview plays the music as the render will mix it.

How loud the music really plays is measured, whatever its source: after each save (as part of the quality check) and again on the render. Three numbers are given, on the scale of the finished video: its loudness where no one speaks, its loudness under the voice once lowered and how far below the voice that is, and its longest stretch quieter than -35 LUFS. A track that is silent or too quiet to hear, a quiet passage of six seconds or more, or music lost under the voice (more than 28 dB below it) is reported with what to change - raise Volume, lower it less under the voice, start the track elsewhere or choose another one. It is advice: it never stops a render.

The clips' own sound

AI shots often come with their own sound - effects, ambience, sometimes music of their own - and an uploaded video carries its own audio. Under the video's music, an AI shot's own sound is off by default, because its music would fight the bed; an uploaded video keeps its sound. Two places change it:

Video > Music and sound > The clips' own sound: Automatic (the rule above), Keep every clip's sound, or Off: only the voice and the music - the music then carries the whole video under the narration.
A clip scene's own Its own sound in the inspector: Automatic (the video's choice), Keep it or Off. The inspector says what the automatic choice does for that scene, and the timeline marks every scene whose own sound plays with a speaker.

The assistant does the same with clip_sound on the video and audio on a scene, and tells you what plays where after each change, naming the shots that have no sound at all. An AI shot made for a video whose clips' own sound is Off is made without sound: the track the model adds anyway is dropped as soon as the shot arrives.

The voice always leads. A clip's sound that plays is levelled toward the voice (a loud shot brought down, a quiet ambience left quiet), set by its Volume against the voice (0.8 by default), faded in and out over the transitions around it, and lowered by 10 dB under every sentence, like the music. Under a narration the voice sits on top, the clip's sound under it, the music under both.

The final soundtrack is normalised to -16 LUFS, the loudness streaming platforms expect for speech-led video.

Captions are drawn on the video (or not: Burn them into the video) and always delivered as a WebVTT file next to it.

Media and their retention

The Media tab of the Video Studio lists your recordings and AI shots together, newest first and a page at a time, with the videos that use each one; the page bar sits above and below the grid, like on the Videos tab. Search them by name or prompt, and filter them by kind, by use (In a video or In no video) and by when they were made (a preset such as Last 7 days, or a From / To day): the search and the filters cover your whole library, not only the page shown. "In no video" means no video at all uses it, a colleague's included: a card used only by a video you cannot open says In a video you cannot open.

To clean the library up, select media and delete them together: tick cards one by one (across pages), tick the box of the bar above the grid for the whole page, then select all for every medium the filters match. Select all unused does it in one click for every recording and AI shot no video uses. A medium counts as used when any video needs it: a scene plays it, a screen insert, a backdrop or an image shows it, an AI shot the video plays was made from it (its first frame, the recording or shot that frame came from, its reference pictures), or a ready render of the video drew it. A deletion of several media never deletes one a video uses: those are kept, the confirmation and the result say how many, and deleting one of them anyway is done from its own card, which names the videos and asks you to type a code back. The use is checked again as the deletion runs, so a medium placed in a video meanwhile is kept, and only your own media are ever deleted.

A recording or a shot that failed or was cancelled made nothing: it is not listed, and once no video uses it, it is deleted an hour after it failed. Their footage is the heaviest part of the studio, so the footage no video uses is kept for a while, not forever:

the media a video uses are never deleted, however long the video is left untouched: it can always be edited and rendered again;
once no video uses a recording or a shot any more (removed from the timeline, or the video deleted), its media are deleted some time later (30 days by default);
the media of one no video ever used are deleted some time after they were made (14 days by default);
a rendered video keeps its own MP4: deleting a recording's footage never touches a video already rendered from it;
the card of a recording or a shot no video uses says when its media expire; Keep its media exempts a recording or a shot;
an expired recording or shot stays listed, marked Media expired, with its name, poster and script, so it can be recorded or generated again; a video that still uses it can no longer be rendered until the scene is replaced - or, for an AI shot, until Generate again in the inspector makes it again in place, for every video using it.

Deleting a recording or a shot from the Media tab removes it and its footage for good. One that no video uses is deleted after a plain confirmation. One that a video uses is protected: the confirmation names the videos that use it (and counts the ones you cannot open), and Delete anyway stays disabled until you type back the code it shows. The API refuses such a deletion too (409 media_in_use) unless it is forced.

If a medium a video uses is deleted anyway, the video keeps opening, previewing and saving: the scene that played it shows This recording was deleted (or This AI shot was deleted) where it plays, and the rest of the video plays around it. Record it again or give the scene another recording or shot in the inspector; until then the video cannot be rendered, and the render names the scene.

Administrators set both periods in the deployment configuration (see Video Studio settings).

Render and watch

Render at the quality you need from the editor's Renders tab: Draft (720p, 30 fps) to check the cut, High (1080p, 60 fps, the default) or Ultra (4K, 60 fps). A render takes a few minutes and continues if you leave the page; keep editing meanwhile, the render works from the revision saved when it started. A video has one render in progress at a time. Asking again for a render of the same saved version at the same quality - from the editor, another tab or the assistant - follows the render already waiting or running instead of starting over, and the history shows it once. A render of a newer version or at another quality cancels the one still waiting or running, which stays in the history marked Replaced by a newer render - except when the newer version changes nothing the render draws (a screen insert's legibility, a deck's speaker notes, a clip's own sound or its level): saving it keeps the render in progress, which finishes as the render of the newer version, its sound mixed for that version, and rendering that version at the same quality follows it rather than starting over. A render that stops moving for two minutes is restarted where it stood, the frames it had already made kept, and its progress says so; after two such restarts it fails. When the platform is updated or restarted, the worker stops taking new renders and lets a render past half its frames finish, for up to 20 minutes where the deployment gives its workers that long to stop (VIDEO_STUDIO_RENDER_DRAIN_SECONDS); a render under half its frames goes to a running worker at once. One still under way is handed to the next worker, first in the queue, which resumes it from the frames already made and says whether it resumed after an update or a restart: an update never fails a render. A render handed over keeps its place on the renderers: a running worker keeps a slot free for it, and when the renderers' slots were all taken meanwhile by renders queued after it, the one that started last gives its slot back - keeping its frames and its place in the queue - so the render handed over resumes within a minute. A worker that dies without stopping (out of memory, a crash) is noticed within seconds, its render taken back first in the queue and resumed, and the render says how the worker went - the signal and the memory it used then. The render list and the assistant say how long a render left by a stopped worker has waited and how long it may still wait: as long as an update of the platform can take to replace its worker - 10 minutes unless the deployment says otherwise (WORKER_TERMINATION_GRACE_SECONDS, WORKER_STARTUP_SECONDS) -, and for as long as another render holds the slot it waits for. A scene preview holds that slot only when it runs in a render slot lent to it, and gives it back to the waiting render, which takes it first; previews on their own slots are listed apart and never make it wait. Once its slot frees, the render still has its whole wait before it fails. One that no worker resumes by then fails instead of showing as running, saying which workers were running and why none took it. Render it again at the same quality within a day: the same render is queued again first and resumes from the frames it had already made, instead of starting over. Before it starts, the assistant says how long the render should take and what that estimate is made of: the frames, the seconds a frame takes on each page of the renderer and where that figure comes from (the platform's past renders at that quality, its renders at other qualities, or a reference measurement when it has none), the pages, the CPUs and the memory. The frames are counted by what they show - one scene, two scenes during a transition, or a screen insert, which costs a page many times a plain frame - and by the type of scene they belong to - a clip with a screen insert, a statement over a clip, a figure -, each at the time a page took for a frame of that scene type on the platform's past renders and scene previews (by its kind only for a type never drawn here; a screen insert's frames at the time the previous render of the same video took for them), with the time the screen inserts' preparation and the checks of the finished video take, and the narrations at the length they were spoken when they already were. The estimate is quoted from the slowest scene type the video holds: how many frames it has and what one takes a page, since the page drawing those scenes decides how long the render takes. It says how much one page of the renderer holds at its peak (measured on a 4K page drawing screen inserts, for Ultra), how many pages the renderer's memory opens, and how long the render takes on them against on every page, with the memory to give for every page. It also says what its last estimates at that quality said against what those renders took, and corrects itself by that. It counts only what the render will draw: an animated figure an earlier render of the video already drew, unchanged - at that quality or sharper - is reused rather than drawn again, and the estimate says how many. The renderer's pages share the frames by what they cost, and a page done with its share takes frames from the pages with the most left, so the frames of a long screen insert are drawn by every page rather than by the one or two that started them. While the frames are drawn, the render's progress says how many frames are left, how many it drew a second over the last minute and how long the rest should take. Once drawn, the assistant says how many frames the render drew and how many it kept from an earlier attempt, so a render that resumed after an update is never mistaken for a faster one. A render one renderer cannot finish within 4 hours - or, when the platform runs several renderers, within 30 minutes on the pages its memory opens (VIDEO_STUDIO_RENDER_SPLIT_MINUTES) - is cut into scene ranges, each drawn by a renderer of its own, then joined with its sound once: with several renderers, a long Ultra video renders in a fraction of the time. A range that no other renderer takes is drawn by the render's own renderer after its range, so the render never waits for a renderer the platform does not have. An Ultra render the renderers cannot hold (too little memory, too slow even split into ranges) is refused until the platform has rendered at that quality once. The refusal says how long the render would take, the CPUs or memory a renderer needs, and the highest quality that fits now, for you to choose. You can also render anyway. A second render of the same version at another quality reuses what does not depend on the quality: the narrations, the figures drawn by a sharper render, the music composed for the video and the screen inserts followed through their footage. The result is an MP4 (H.264 video, AAC sound) with a poster image and the captions file. The poster is a settled frame of the first scene after the title card, never a frame caught in the middle of a transition. When the captions are burned into the video, the player leaves the captions file off, so they never show twice; when they are not, the player shows the file in the band the scenes keep clear for it.

To check one scene you changed without rendering the whole video, ask the assistant to preview it ("preview the booth scene"): it renders just that scene with its voice, its music and its transitions in and out, usually in under a minute, and shows you the video and its first, middle and last frames - the scene settled, out of its transitions and of the video's fade from and to black, an opening or closing card your template adds included. A scene showing a figure, a recording or a screen insert is previewed at 1080p (30 fps), so you judge its sharpness on what a render shows; any other scene is a 720p draft (ask for a 1080p preview of it if you want one). A player scales a preview to its screen, so a 720p draft looks softer on a large display than the render will: the assistant says the size each preview was drawn at, and the editor's Renders tab lists the video's scene previews with that size under each one, to play them full screen. Scene previews run beside the whole renders, on worker slots of their own: a preview never waits for a 4K render, and a render never waits for anyone's previews. The worker keeps memory for its previews that no render takes, so a preview always draws at its full speed: when its worker is already busy with other previews beside a render, it goes to another worker with room, or waits for the others to end, and its answer says which. While a worker has no whole render to make and its preview slots are all busy, its render slot takes the next preview too: a whole render asked meanwhile waits for that one preview to end. A preview never replaces the video's current render, and is kept for a day. Ask for as many previews as you like - every scene of the video at once if you want: each one is queued, a few of one video run at a time, and each free slot takes the videos in turn, one preview each - the video with the fewest previews being made first, then the one served longest ago - so a sweep of twenty scenes never waits behind every preview of another film, and the preview of a video asked after another's whole sweep comes next. A preview waiting for one worker's memory is taken at once by another worker whose slot frees, in its turn. While a preview waits, the assistant names what it waits behind: "1 preview of video Launch film ahead", the previews being made included, and how many workers draw previews ("1 worker draws previews, 2 at a time"). A preview that waited behind other videos' previews on a platform with a single preview worker says once, for the administrator, that a second preview-only worker would start another video's preview at once. An edit of only the video's theme (its watermark, its backdrop, its colours) or its music keeps the preview every scene passed: no scene needs a new preview for it, and the whole render draws and checks the new look. A figure is drawn once and kept with the video: the previews of the scenes around it, and a second preview of its own scene, reuse it rather than drawing it again, and a scene beside a figure draws only the moment of the figure its transition shows. A screen insert on a scene before or after the one previewed is drawn as in that scene's own preview; if its source cannot be read, the preview fails saying which, rather than showing the scene without it - and so it does for a missing backdrop or AI shot of a scene beside it, with the reason that scene's own preview would give. The assistant says where a preview stands in its lane of the queue, and behind which videos' previews. A preview is made exactly as a whole render makes that scene, and gets the same sound and picture checks (below), so you can validate a scene before rendering the whole video. The narrations of the scenes before and after it play in its transitions and are heard back too: one that fails is a warning on that neighbouring scene, never a failure of the scene previewed. So is what the picture check finds on a neighbour's screen insert shown through a transition: it stays on the scene that owns the insert, a warning that names it (or left out, when that scene's own preview already reported it), never a failure of the scene previewed.

Every render is heard and watched back

Each narration is heard back before it is placed. Once spoken, a narration is transcribed and compared with its text. It fails when more than one word in ten differs, or when the speech repeats words the text does not. A name with no pronunciation entry (a capitalized word such as Moscone Center) counts as heard when what was transcribed sounds like it, since speech-to-text spells a name the way it heard it. A narration that fails is first given the take of the same voice that already passed on the same words, for example in the scene's preview. Without one, it is spoken again with the same voice, then once more a little slower (85% of the video's pace) with the pronunciations written into its text, then with the voice of the same speech model closest to the video's - the one sharing the most of what the provider says of its voices (accent, gender, tone...), never another provider's or the platform's default voice. A narration that fails every time is never placed silently: the render stops and says which scene and why. A narration that passed is not checked again in later renders.
Another voice is never a silent pass. When only that closest voice says a narration right, the audience would notice a different voice in the middle of the video. The render plays it and lists a warning for that scene (Narration spoken by another voice) naming the voice used, the video's voice and why it failed; the quality check shows it too, as a finding the audience would notice. Until the video's own voice passes, Render refuses the video for that scene. Reword the narration or add a pronunciation, then preview the scene to try the video's own voice again: once it passes, the video renders as usual. To keep the other voice, make it the video's voice, or render anyway: the video is rendered with it, and the finding stays on the render.
A scene preview checks the narration the whole video uses. A preview speaks each narration exactly as the whole video does, with the narrations around it, so a narration that passed in a preview is placed as it is in the next render, without being spoken or checked again.
The voice-over has its own clean track. Every narration is placed at its scene's start, to the sample, with a short fade at each end. Two narrations never overlap: a scene lasts as long as its narration.
The finished video is checked. Once the video is assembled, its soundtrack is played back and compared with the voice-over track. The render fails when more than one word in ten of the narration is heard wrong, when a sentence is heard twice, or when the same fraction of a second of sound repeats three times in a row (a stutter). The picture is watched too: any stretch of over half a second where the frame does not change is reported - a locked-off shot where only a small subject moves (a cat far down a corridor), even one barely lighter than the wall, is not frozen while it moves; a shot carrying a screen insert is watched in parts, its footage and each insert apart, so a static page on a screen in a room where a candle flickers, or a still room under a page that scrolls, is not frozen, and a frozen picture says which part stopped moving - except where stillness is the design - a crossfade or a fade through black, where the picture holds while it fades, title cards, closing cards and statements, an image or a slide set to no motion, an animated figure resting between its steps, a recording while the product it shows is at rest (its held end, after the recording, is checked), and your own Hold the last frame at the end of a scene. A frozen picture is a warning: the render is ready. The music is measured as it was mixed (see "Music" above): its levels are shown under the render in Renders, and a quiet mix is a warning too. While the finished video is checked, its progress says how many narrations have been heard and how far into the picture the check has watched, so a long check is never taken for a stuck one.
Every finding has a time. Each finding says where it is in the video (1:02.35) and in which scene. They are listed under the render in Renders and in the assistant's answers. A render that failed its sound check keeps its video so you can play it at those times; it never becomes the video's current render.
A render's answer starts with its verdict. The assistant's answer for a render or a scene preview says first whether it is green (nothing blocks or warns) or blocked. It then lists, apart, the findings (each with its time and what to change), what a setting you chose shows by design (the bands a screen insert set to Fit inside the frame leaves on a wider surface, its colours at the light you gave it), and the measurements that block nothing. Only findings are issues: a render whose findings are empty passes every gate. Advice on sizing the platform's workers is for the administrator, and is said once, in the first render that needs it.

Speech-to-text uses the same OpenAI provider as dictation in the chat: the narrations' audio is sent to it to be transcribed, whichever provider speaks them, within the providers each person may use. Each narration is transcribed once: the check of the finished video reuses what was heard when the narration was checked, as long as the encoded voice still plays that very take, and transcribes again only a narration the stitching or the encoding changed, or one checked before this was kept. The render's checks name the speech-to-text model and provider (the provider's name as set in Settings > AI Models). Without one, the narrations are placed unchecked and the render says so. The stutter and frozen-picture checks still run.

A clip, a recording or an animated image shorter than its narration keeps moving to the end of its scene by default: it plays slower to fill it, then drifts in on its last frame (see "Footage keeps moving to the end of its scene" above). A scene's When the narration outlasts it setting can instead play it at its own speed and then loop its last second with a crossfade (clips and animated images only), or play it at its own speed and freeze its last frame. A frozen last frame is reported when you save: the quality check compares the clip's or the recording's length (its trim, its speed) with its narration - at the length it was spoken, or at the reading pace until it is - and a scene set to freeze whose last frame would stay still on screen for more than half a second is a finding that stops the render until you change it, suggesting the drift or the loop instead, so no render is paid for to report it.

The render checks before it starts. Render refuses, with its reasons, when an AI shot or a recording the video plays is still being made, failed or expired (each with what a render would do with it), when the quality check saved for the video finds something the audience would see (advice alone never stops it), when a narration failed its last check, when the previous render of the same revision failed its sound check (the same revision would sound the same), when the previous render at the same quality showed a frozen picture (a draft that showed one never stops the final render), or when a scene changed since it was last previewed or rendered - its words, what its preview shows of the scene before or after it (their picture and the transitions between them, not their narration), the video's look or its music, its transition (its own, or the video's default when it has none), whether a clip keeps its own sound, the video's fade on its first or last scene, what a pronunciation applied to the captions shows, or a new take of the recording or AI shot it shows. Editing what one scene shows marks at most three scenes as changed (it and its neighbours), rewording its narration only that scene (unless it changes a transition beside it); the others keep their preview. What counts is what a scene shows, not when it was saved: put a scene back as it was, even after previewing the change, and the preview made of it before counts again (the save says so). A save that changes nothing at all is not saved: no new revision, nothing to preview again. A scene nobody previewed yet never holds a render back, and neither does a change of voice, delivery, pace or of how the voice says a term: the render hears every narration again before it draws anything, and stops on one that fails. A preview that passes is trusted: a finding about an AI shot (its camera moved, for instance) no longer holds the render once every scene showing that shot was previewed as it is now and passed, nor does what an earlier render found in a scene previewed since, so a video whose every scene preview passed renders without being forced. You can render anyway: the choice is recorded on the render and in the audit log.

The assistant renders a video once it is ready. Refused, the assistant fixes what it can, waits for the shots, previews the changed scenes all at once (the previews run side by side) and looks at their frames, then renders. Ask it to render anyway and it does (force).

Renders wait in a queue for a free renderer, and each one says where it stands (being made, how many ahead of it). One person can have a limited number of renders queued or running at once, across all their videos (10 by default, VIDEO_STUDIO_MAX_RENDERS_PER_USER); past it, the render is refused with the ones in flight listed, so you can wait for one or cancel one (from its video's Renders tab, or by asking the assistant, which can also list everything of yours still being made).

When the assistant starts a render - or a recording, an AI shot or an AI track - that is still running when its answer ends, it comes back to the conversation once the work is done, without being asked: it shows the rendered video, says what failed and why, and carries on with your request if a step of it remains. When it started several at once, it comes back once, when the last of them ends. Stopping the conversation keeps it quiet: the work still finishes, and the assistant says nothing until you write again.

Each render runs on its own, within a memory budget the administrator sets (VIDEO_STUDIO_RENDER_MEMORY_MB), so a render that fails never stops anything else the platform is doing. A render that crashes, or whose renderer stopped answering for a moment, starts again by itself at the same quality, up to three attempts in all, and picks up where it stopped: the frames and the animated figures it had finished are kept, and only the rest is drawn. Frames are kept only when everything they were drawn from is the same - each scene, each file, each screen insert's every setting down to its key's tolerance and feather -, so a video edited meanwhile is drawn again, and the render's answer says which frames it kept, for which saved version and by which attempt (or that it kept none: a preview or a render never takes another render's frames). A render that stops moving forward for two minutes is resumed the same way, twice at most. A render that is still drawing is never failed because its renderer checked in late: it starts again from the frames it had finished, even after its last attempt. When it fails for good, or uses more memory than its budget, it says why: how far it got, which scenes it was drawing and how much memory it used. The video and its quality need no change: the way through is more memory for the worker, which the administrator gives. A failure of the platform rather than of the video - its headless browser failing, its memory, a crash, a stall, its worker going silent - first runs the render again once, by itself, while the video has not changed since, keeping the frames it had finished. A render that still fails says the failure was the platform's, that it already ran again once, and that nothing in the video needs to change. A failure of the platform rather than of the video - its headless browser failing, its memory, a crash, a stall, its worker going silent - first runs the render again once, by itself, while the video has not changed since, keeping the frames it had finished. A render that still fails says the failure was the platform's, that it already ran again once, and that nothing in the video needs to change. While a render runs, the assistant can see when its renderer last checked in, and says so when it is late.

Each render that finishes becomes the video's current render: the one the library card, the chat card and the public link play. To play an earlier one instead (the 1080p cut rather than the draft you rendered after it), open its menu in Renders and choose Make it the current render.

Rendering changes what the video plays, so it is for the video's editors. Someone the video is shared with as a viewer duplicates it to render a version of their own.

Watch a video

Play opens the video over the whole window, on black, and in your browser's full screen at once - never in a small box. It is in the editor's header (the current render), on each render in Renders, on a rendered video's card in the library (the play button on its picture, or Play in its menu), and on every recording and AI shot in Media. The title, what is playing and Download sit at the top; Esc leaves the browser's full screen, and Esc again, or Close, takes you back where you were.

The player is the same everywhere a video plays - there, on a viewer's page, on the video's card in a conversation and on its public link: play and pause, 10 seconds back and forward, a timeline that shows what is loaded and the time under the pointer, the volume (remembered from one video to the next), the captions on or off when they are not burned into the picture, the speed (0.5x to 2x), loop, picture-in-picture and full screen. While the video plays, the controls and the pointer step aside after a moment; move the pointer, or tap the video on a phone, to bring them back. Click the picture to play or pause, double-click it for full screen; on a phone or a tablet, double-tap its left or right side to go 10 seconds back or forward (tap again to go further).

Chapters. A render made since chapters exist has one per scene: the timeline is cut where each scene takes over, the pointer over it names the scene, and the chapters button next to the time names the one playing and lists them all with where they start, to jump to one. A scene is named as on the editor's timeline - a card's title, a statement's first line, the recording's or the AI shot's name - else by its place and type (Scene 3 - AI shot). Your keyboard's media keys and your phone's lock screen play, pause and skip the video too, and their previous / next track keys move between chapters.
Where you left it. Leave a video before its end and it starts where you stopped the next time you play it in this browser, with Start over for a few seconds. A video watched to its end starts from the beginning.
When the connection drops. The player picks the video up where it was once it can; if it really cannot play it, it says so with Try again.

Press ? in the player for the list of its keys; it is also in the menu behind the speed (1x), with Loop.

Key	Does
`Space` / `K`	Play / pause
`Left` / `Right`	5 seconds back / forward
`J` / `L`	10 seconds back / forward
`Up` / `Down`	Volume up / down
`M`	Mute / unmute
`C`	Captions on / off
`F`	Full screen
`0` to `9`	Jump to that tenth of the video
`Home` / `End`	To the start / the end
`<` / `>`	Slower / faster
`,` / `.`	Previous / next frame (the video pauses)
`?`	The list of the player's keys
`Esc`	Leave full screen, then close the player

Download

Download, at the top right of the editor, saves the current render: the MP4 (its quality and size are shown), the captions as a WebVTT file and the poster image. The files are named after the video and its quality, for example Launch film - 1080p60.mp4. The same entries are in the menu of the video's card in the library and on its card in a conversation, and every render in Renders downloads its own files. Until the first render is ready, Download says so.

Share, at the top right of the editor (or in the menu of the video's card in the library), works like the presentation one:

People and groups. Give access to people with an account here, or to groups, as Can view (watch, download and duplicate) or Can edit (also edit, render and share, in the editor or through their own assistant).
The company. An administrator can publish the video to everyone in the company; only administrators then change it.
The link to copy. The video's address opens it for the people it is shared with, and nobody else: the dialog says who it opens for before you send it. Copy link in the editor's menu copies it too.

Ask your assistant to do it - "share this video with ada@example.com as an editor" - and it changes the sharing with the dialog's rules: only the owner or an editor shares, and only with existing accounts. Every change, from the dialog or the assistant, is in the audit log with who gained access, who lost it and whose access changed.

The library lists your videos, those shared with you and those of the company, the last changed first and a page at a time, with a filter for each and a search that covers the whole library (title, template or owner). The page, the filter and the search stay in the address, so a link or a reload keeps them. A globe marks a video with a public link, an eye one you can only view.

Public link

For people who have no account here - a customer, a partner, an event audience - turn on the video's public link in the Share dialog. Anyone who has the link watches the video's current render in their browser, without signing in, in the whole window with the same player (see Watch a video): never its scenes, recordings, AI shots or editor, and nothing about who made it or who it is shared with. A new render, or another render made current, is what the link plays next.

Copy public link copies it; Replace link issues a new one, and the old one stops working at once.
Allow download (off on a new link) gives its visitors a download button for the MP4 and its captions. It is a convenience, not a protection: anyone who can watch the video can save what the player plays, with the option off as well.
Turning the link off makes it answer "not available" at once; the people and groups the video is shared with keep their access.
Until the video is rendered, the link says the video is not ready yet.
One address can pull at most 2 GiB of the video's files per minute - some thirty viewers of a 1080p60 render behind one office address - past which the link answers "too many requests" until the minute is over.

Your assistant can turn the link on or off and return it: "give me a public link to this video". A public link is a capability: whoever has the address can watch the video, and can pass it on, so turn it off or replace it when it is no longer needed. Administrators can turn public video links off for the whole platform in Settings > Security > Security options (Public video links, or VIDEO_PUBLIC_LINKS_ENABLED=false, see Security options): every link then stops working at once and none can be created; turning the option back on restores the links that were not turned off.

Render retention

Every render keeps its own MP4, poster and captions, so a video's renders are kept for a while, not forever:

the 5 newest renders of a video are kept, each 30 days after it finished (failed attempts are counted apart, so they never push out a render that worked);
the current render is never deleted, nor a render in progress;
Keep it, in a render's menu, exempts that render; Delete removes one at once (deleting the current render makes the newest other one current).

Each render in Renders says until when it is kept. Administrators set both rules in the deployment configuration (see Video Studio settings).

Administrators find every video, whoever made it, under Administration > Objects (filter Videos), see Objects. A video posted in a conversation on another channel (Slack, Teams, email) is a link to its page.

Limits and settings

Administrators set the recording and video length limits, the number of renders that run at once, the media and render retention and the ffmpeg location in the deployment configuration (see Video Studio settings).