SHIPPED · v0.11.0
Edit your own footage from the transcript, on a timeline you can cut
New
A timeline you can edit. The Timeline tab now has video, audio and title tracks with a live preview. Drag, trim and split clips, scrub or step through frames, and use snapping and undo/redo to refine the cut.
Make the cut your own. Import pictures, video and audio, add titles, adjust their size, colour and position, and set the canvas background. Rename, lock, hide or mute tracks, fit the timeline or expand the editor, save the cut with the production, and export it as an MP4.
A calm screen-demo look. Screen recordings get a dedicated 16:9 recipe that eases in on the parts of the interface that change, with a light music bed and no flashes or shakes.
Edit your recordings, not just generate them. A new Edit footage workspace sits next to the prompt composer on the Hub and on the rail. Drop in video or audio recordings, pick a profile — Short vertical, Conversational, Essay or Story — and the shapes you want to deliver, then press Analyze. The app transcribes the speech, measures the pacing and finds the faces on this machine, then proposes every change as a suggestion you can accept, reject or lock.
The transcript is the editing surface. Cuts, shortened pauses, caption keywords, punch-ins, take picks and voice patches are marked right in the text. Select words and press Delete to write your own cut; click any mark to open the suggestion behind it — who proposed it, its confidence and the numbers it measured. Group actions such as "accept all pacing above 0.85" undo as one step.
Pacing, takes and hooks. The cut list trims pauses and fillers to the profile's rhythm. Repeated attempts at the same passage group into takes with the strongest one picked — and flagged when the repetition looks intentional. A hook extractor offers the strongest lines to open the video on, with a guard against spoiling the ending.
Word captions, punch-ins and reframing. Animated word-by-word captions in five bundled styles with keyword colour, zooms that land on emphasis, and a landscape recording can come out vertical, cropped to follow the active speaker — low-confidence spans are flagged rather than silently guessed.
B-roll, music and chapters. Cover a phrase with a picture from your library or a generated one, pick a music bed that follows the speech's mood and ducks under the voice automatically, and get chapter titles with timestamps ready to paste into a description.
Fix a misspoken word. Retype the word in the transcript and a voice patch reads it in the speaker's own voice — only after you attest the voice is yours or permitted, hear the patch and accept it. Patched words, generated pictures and corrected eyes are disclosed in the delivery notes.
See where every job runs. Each step carries a mark — on this machine, leaves this machine, or off — computed from what it will actually do. A new Footage analysis tab in Settings names the engine and model behind each job, and before transcript text first goes to your coding agent the app shows exactly what is sent and what never is.
Your manual tweaks survive re-renders. Adjustments you make to a generated film are kept as layer overrides and applied again after every re-fill or re-direct; an override that no longer fits is reported, never dropped.
A look menu that leads with Auto. The format menu and decision chip are now one control: it opens on Auto's pick with its swatch, reason and cost, plus three pictured alternatives, and Kind, Frame and Length as small fields you can pin.
A gallery of all 30 looks. Browse every look as a picture, filter by what the film is for, by kind or by what this machine makes well, and search by name, description or reference. Picking a look your kind doesn't offer moves the film to a kind that does, and each style has its own hand-drawn swatch.
Improved
Travel edits follow the trip. The travel diary recipe uses capture dates to order photos and video moments, adds gentle photo moves and dissolves, and keeps its music lighter under the recordings' own sound.
Footage fills the chosen frame. Recipe crops use measured subject tracks, salient regions or faces, and overlays keep clear of faces.
Saved timeline cuts keep their timing. Generated scene media can refresh without replacing your manual cut. Video trims, title timing and track audio settings carry through to the exported MP4.
Stricter delivery checks. A complete video decode is required, and failed or missing media measurements are reported as failed checks. Render feedback can repair actionable picture and final-QA findings while retaining the evidence from earlier attempts.
Auto picks with evidence. Look descriptions and reference names are read in English and Vietnamese, and every offered style is scored on your words, use tags, attachments, cost and your recent films — with the confidence, top scores and next three looks shown. Styles that need image or video generation, photos, footage or Blender are only picked unasked when the run can supply them.
Auto asks when it guesses. When a pick is a guess and nothing is pinned, the Brief stage asks once which of the top five looks you'd prefer. Your answer is recorded on the decision and said in the chat, and the rules pick stays available.
A look you name is honoured. Pin or name a look and the run keeps it — even when some of its effects need simpler fallbacks, which is noted rather than swapped silently. A hand-picked look also survives the variety pass.
3D stays 3D, with a warning. A trailer, brand film or music video that goes over the render budget stays a 3D film with a note on the render time; it only moves to a 2D recipe when the prompt reads 2D or 3D is unavailable. Fiction reads as Story in the pick.
Styles follow their own recipes. The director now gets each style's recipe card — its frame rate, posing, drawing speed and flash limits — the shot plan is checked against it, and rendered reels are linted for the look's own rules.
Sane export sizes. Rendered videos carry a bitrate ceiling that scales with frame size and rate, so noisy footage can't balloon the file, while small frames keep a usable floor.
Fixed
Fixed edited images, videos and same-named palettes showing stale pixels in Ink previews.
Fixed clips, masks, effects and group opacity around live video or Ink layers losing their context during Composition rendering.
Fixed caption selection order disagreeing with the visible overlay order, and reject invalid export ranges before touching the output.
Fixed 3D preview passes changing final motion blur, exposure settings or certified character placement. Resume checks include the actual calibrated render settings.
Fixed malformed Whisper Unicode output failing the whole transcript, and noise-floor checks flagging intentional jumps between recipe shots.
Fixed a compatible local narrator being replaced during language routing, and bounded stalled macOS speech-synthesis jobs.
Fixed "the 90s" being read as a video length — naming a decade in the prompt no longer sets the duration.
Fixed the composer showing paid generation lanes as usable when advanced generation was off — the pick now matches what the engine would actually do.