Changelog

See what's new in Lucy

0.41.0 — 2026-08-23

Share a project, and start from somebody else's shot

A link that hands somebody a read-only copy of a film, forty-six reviewed prompts to start one from, and a single score written from the plan instead of eleven unrelated pieces of music.

Features:

A project can actually be shared with somebody. Asked plainly — "so how do I share the project with another team member?" — and the answer until now was that you could not, which was not obvious: the server has been able to mint a token, grant a viewer role and clone a project since the day before. What was missing was any way to ask for one.
A film can start from somebody else's shot. Forty-six reviewed prompts across ten categories, from a CC BY 4.0 library, each one carrying its author and a link to the original — because the credit travelling with the text is the whole of what that licence asks.
The importer refuses what a review would have removed. Screening by hand only works until the next scheduled import puts back what somebody took out that morning — with nobody watching the job that did it.
Every project and every film gets a working title. "New project", "New project", "New project" is not a library, and neither is three films sharing the first line of a brief.
Lucy writes the music brief from the film, and one score is written from the plan before the first take. Music was the last thing in the product a person had to describe by hand — fine when you know what you want, useless when the question is what music this film should have, which the plan already answers. Every take used to be generated on its own, so an eleven-shot film carried eleven unrelated pieces of music cut end to end. That is a collage, not a score.

Improvements:

The template picker shows you the prompt before it replaces your brief. It listed forty-six titles with a Use button on each and not one word of what any of them contained, while Use throws away the brief and replaces it with up to nine thousand characters you have never seen. It is two panes now: the list narrows by search and category, and the button that replaces the brief sits with the text it is about to paste.
Every visit starts on the work you have already made. The landing screen used to be chosen by whether a brand existed — the month if it did, an eight-field intake if it did not — and both are places to do work rather than places to find it.

Fixes:

Films made from a phone are no longer silent. Every one of them was mute, measured on a finished film rather than inferred: a take whose dialogue reads "Battery acid with a hint of regret" had exactly one stream in it.
The voice step cannot block a film that has no voice step. Nine shots, seven approved, stills already paid for, stopped dead with nothing in flight and no failure recorded.
Pressing Cut on a finished film stops instead of spinning forever. Reported as "I click on it but it kept spinning and I don't know what is happening" — the film had been fine the whole time, 43.7 seconds and playable. Only the answer was wrong.
Two boot hooks that a rewrite silently took with it are back.

Changed:

Lucy stops wearing Telegram's clothes. Lucy-branded rather than Telegram-native, cinema school rather than call sheet, and the moment a cut arrives treated as the one thing worth remembering.

0.40.0 — 2026-08-22

Projects stop living in one browser

A project now exists on the server rather than only on the device that made it — so it opens on a second machine, survives clearing site data, and every charge that ends in a failure comes back.

Features:

A project is a row on the server and its footage is in a bucket. Until now projects were IndexedDB and OPFS: a film existed on the device that made it and nowhere else, so clearing site data lost it and a second device could not open it at all.
A shared project is read-only until you clone it. The clone points at the same asset objects, so a hundred forks of one film cost one film's storage.
An upload that usually does not happen. The key is the content hash, so the question before sending anything is whether it is already there — and for a clone, a re-import after a browser was cleared, or two people with the same footage, it is.
The dashboard is told about any change as it happens, over server-sent events rather than by asking repeatedly.
The mini app plans a month before it films anything. It used to open on "describe your film", which is stage five of the workflow — no company, no month, and nothing anyone could approve.
Filming carries on when the phone is gone.
One tap past the voice gate. The dead end it produced was reported from a full production run; the message was in fact correct, and the thing it asked for was the part that could not be done on a phone.
A 500 says which one it was. Every failure now mints a short id, logs it, and writes it to a table that survives the next deploy — because the last outage was undiagnosable, and that was the fault worth fixing rather than the bug that triggered it.
A build refuses to serve a database it cannot write to. The drift check was a command somebody had to remember to run, which is the same category of protection as a sticky note; it is a release gate now.

Improvements:

1080P on Seedance holds fifteen seconds, measured twice at 1920x1080 and 15.04s. The old ceiling was five because five was all anyone had tested, and the comment in the code said so.

Fixes:

Every charge that is followed by a failure comes back. Measured on the Telegram route as two ten-token deductions for one delivered clip, and the same hole was in every other charging path — they all bill the provider first and can still throw afterwards.
A generation that cannot start gives the tokens back.
Cloud saving worked exactly once.
Two different clocks were being compared, and a delete did not stick.
A rename never reached the cloud.
A document with no scenes stops bricking the project, and the server stops accepting documents that are not projects in the first place. Two projects created minutes apart opened to "No active scene." and stayed that way across every reload.
The resolution shown is the resolution that comes back.
A request too heavy for the model steps down instead of being bought and then refused.
The drafter writes the captions it was always meant to. Measured across six real plans before anything was changed: two captions in thirty-one shots, and a film drafted as three title cards came back with all three cards blank.
A finished film stops re-cutting itself every six seconds. It was not stuck because a cut had failed — it already had a finished film — it was stuck because it kept cutting.
A new storyboard stops being filmed as the previous film.
The bot says why it failed, instead of one sentence that fits every cause.
A post learns it is finished from the film rather than from a phone, and a post's angle is the point of the post rather than a description of it.

Technical details:

Four quota tests that were testing nothing. Reported as ordering flake, and it was worse than flake — they passed regardless of what the code did.

0.39.0 — 2026-08-21

A storyboard you can go through one shot at a time

Approve, redraw or rewrite a single frame instead of the whole film; captions arrive as their own layer; and a mask can say what should happen rather than only what to remove.

Features:

Approve, redraw and rewrite one shot at a time. Every frame carries its own confirm, and the bulk action stays for when you actually mean all of them.
The mini app is a contact sheet, not a scrolled script. Eight full-width cards of prose and an eight-row checklist on a 390px screen put the film — the thing somebody opened the app to see — below the fold, with a list of words about the film above it.
Takes get their own reel, above the panel that counts them, and are cut into a film on the server. A number is a fact about the film; the film is the film. Seven takes and no cut is not a film at all, and there was nowhere else the cut could happen — the mini app has no timeline.
A character can let the model pick its own voice. Casting one is a decision 1,207 options deep, and until now it was the only way to get a mouth that moves with its own audio.
The caption arrives as its own text layer. Asked for because captions born welded into the picture do not look good, and an editor wants the pieces separately.
Drag your own storyboard onto the shots, for when the board already exists — and attach a scene reference from the same library picker the AI tab uses, showing pictures rather than a list of filenames.
A product the brief never named can join the cast.
Room sound is a choice the film makes, rather than whatever the model welded into the take.
Paint over what should change and keep the rest — and say what should happen there, not only what to remove. "Remove object" sent the whole frame to the image model and took back whatever came out, so it did not edit the picture, it re-drew it.
The edit is shown what the answer looks like. A cat-food shot whose cast is a ginger longhair came back with a grey tabby's paw; no model can paint the right cat if it has never seen the right cat.
Qwen3.8 27B joins the drafters as a tier rather than a swap. Proposed as a replacement, which would have moved the whole product off a 235B vision-language model onto a 27B.
Double-clicking an asset drops it at the playhead, and the plus button names both ways onto the timeline. The behaviour was already there and already correct; what was missing was a way to reach it that anyone would find.
A film can carry a music bed, described or brought. Two ways in, for the same reason the cast has two: a described track is fine when nothing particular is wanted, and useless when there is a brand track every film has to carry.
A scene shows its three assets, as the workflow defines a scene — and the AI tab says what it makes.

Improvements:

The assets panel groups by kind, sorts by when you added things, and exports.

Fixes:

Takes come back with sound. Reported as "all generated storyboards seem to be mute" — two separate causes, both ending in silence.
A rewritten scene can be drawn again and confirmed, and so can a shot that is already holding a frame. Reported while stuck: "I can't confirm the scene and I cannot continue."
A film made of nothing but text cards can be assembled.
The assembler's ffmpeg call was malformed and failed on the very first take. ffmpeg was in the container and ran; the arguments were wrong.
Frames are the shape of the film rather than cropped into a guess. QA reported generated images as compressed, and nothing had compressed them.
A cast name the model got slightly wrong still reaches the shot, instead of quietly matching nobody.
The cast voice picker stops loading forever, and the voice popover stops running off the top of the screen — measured on the deployed build at 152px above the window's edge, with no way to scroll to it.
Approving a shot approves it, a chosen voice reaches the film, and the button names what pressing it does. Reported as "I can't approve the character" — the approval had in fact worked, and nothing on screen changed to say so.
The take plays where the copy said it would. "It says tap a shot above to play it but there is no shot to tap" — correct: tapping a frame selected it, and selection took you somewhere else.
A generation that has already landed stops saying "waiting for a slot", and progress counts what is happening rather than what has finished — through a render that takes minutes the old number sat still while the button spun, which reads as stuck.
Stillness stops cutting confetti, and stops freezing the tab. Both found by running the new silent-file support against the 18-minute capture it was built for.
Right-clicking Remove object opens it, and the patch is matched to the frame before it is pasted, so the seam stops showing.
A popover inside a dialog is on top of it, and only the body scrolls.
An assembling marker expires instead of locking the film forever. Writing it before the cut is right until the run dies without returning — a timeout, a restart, a deploy.
The app opens on the project list rather than on whatever this phone last touched, and the page reads in the order the film is made. With one account on several devices, opening the app dropped straight into the casting page of one film while the other handset looked empty.
Smart Clips' suggested moments read in order, and joining them gives you back a cut.

0.38.0 — 2026-08-20

A film can be made on the phone

The Telegram mini app stops being somewhere a film is started and becomes where it is made — cast drawn or photographed, eight steps that say which one you are on, and music under the cut.

Features:

The mini app is the storyboard. A film is made on the phone rather than begun there and handed to a desktop, and a session the bot turns away now says why instead of going quiet.
The phone draws its own cast, and the shot list looks like a shot list.
Eight named steps, so an unfinished film reads as unfinished. Reported as "all the scenes seem empty" — and nothing was wrong: the frames had not been made yet because the cast had no face. The list simply had no way to say that, and eight blank boxes with no explanation reads as a broken app rather than a film at the start of itself.
The phone lists the films it already has. Asked as "should I just generate another project?", where the honest answer was no and the app gave no way to act on it — the film id lived in localStorage and nowhere else, so a cleared webview or a second handset had no handle on work already paid for.
A character can be a photograph you took, and a location a photograph you already have. The references step DREW a face because a webview was assumed to have no file chooser. It has one, and phones do it better than laptops — the picture of the shop, the office, the kitchen is usually the thing a client already owns.
An assembled film comes with music under it.

Improvements:

Upload looks like a button, and generating one says what it costs. It was a chip — a background colour and no border — which on Telegram's own surface reads as a label rather than as something to press.

Fixes:

A film made on a phone outlives the provider's link. Every generated take arrives as a temporary URL that stops resolving; in a browser that is invisible, because the app downloads the bytes seconds after a task lands. A phone does not, so the film quietly expired.
The button offering to open the film in Lucy is gone. It promised a handoff that could not happen — the assets were local to the phone — and it was written before that was true.
The cast kind switch stops reading as a tab. Reported as "when I switch to the Product tab it still shows JIA's name and there is no separate place to upload the product".
A download that times out stops being blamed on your photograph.
The automatic music bed never picks a ShareAlike track. Caught on production an hour after the bed shipped.
Smart Clips says a file is silent instead of quoting Chrome's decoder at you — and then cuts it anyway. Saying so politely was still a refusal, and a screen recording with no sound is a normal thing to want cut.
The film says what it is doing, and shows the frames it has. Reported as "it just spins and then disappears", which is exactly what it did — three faults stacked on top of each other.

0.37.0 — 2026-08-20

Music you can pick, and play without losing the tab

Pressing play with a music bed on the timeline no longer freezes Chrome, adding a track no longer costs 130MB, and the Music tab is a list again rather than a two-row slot.

Improvements:

The tools column opens taller than the timeline. It used to split the window in half, which left the panels that live there — sounds, stock, stickers — about 340px while the timeline sat half empty under a seven-shot film. It now starts at 62/38. The handle still works, and a split you have dragged yourself is kept.
The Pixabay link at the foot of the Music tab takes one line instead of three, giving those three back to the track list.

Fixes:

Pressing play with music on the timeline no longer freezes the tab. The app measures clip length in two units, and the one place that loads a track from the Sounds panel handed back the wrong one — so a 53-second music bed claimed to be 6,360,000 seconds long. Any bed with a fade, which every Fit to project add has, then asked for 382 million volume changes before a single note played. Two quieter faults went with it: a bed added anywhere other than the very start was never played at all, and one added at the start kept playing past its own end.
Adding a track costs a fetch again, not 130MB. Both add paths downloaded the whole file, decoded it to raw samples, and kept them for the session — roughly 130MB per track, on top of a sound card connection that was opened each time and never closed. Chrome allows about six of those per page, so a handful of sound effects quietly broke every later add. Nothing needed either: reopening a project has always worked without them.
The Music tab is a list you can choose from. It had room for about one and a half tracks with the rest behind a scrollbar, which is not enough to compare two pieces of music. The list now takes the space the panel actually has.
One unreadable creation no longer empties your whole inbox. A single row the app could not parse — an edited image, among others — threw away every other creation alongside it. Rows are now read one at a time, so a bad one costs you that row and nothing else.
A sound from Smart Clips can actually be added to the timeline. It was being offered a video track, which will not take audio, so the add silently did nothing.

0.36.0 — 2026-08-19

One voice, one performance, and the controls you could not find

The film stops speaking over itself, the Music tab opens on music, and four settings that already worked are finally reachable.

Features:

Choose the shape of your film: 16:9, 9:16 or 1:1, picked before the storyboard is written. It used to be taken silently from your project canvas, so a square canvas produced a square film with nowhere to say otherwise. Every shot is framed for it, so changing it means redrafting — which is exactly what the brief box does.

Fixes:

Your film no longer speaks its lines twice. Casting a voice after recording your lines left some shots with a recording AND a take that says the same words — two performances layered on one line. Recording is now skipped where the take already speaks, and assembly drops the duplicate rather than stacking it.
Spoken lines sit under their own shot again. A line that ran longer than its shot stretched the picture, and every later line stayed where it was — so a whole film could drift out of sync behind one long read. On the tutorial film that was 3.11 seconds, enough for a line to start while the previous shot was still on screen. Assembly now re-times the lines to the cut.
The Music tab opens on music. It opened on Freesound, which is a sound-effects library — so the music section was full of sirens and heartbeats, with almost nothing over two minutes. Measured on the same search the tab runs: 5 of 20 results over two minutes before, 20 of 20 after. The catalogue was always there; the tab just started in the wrong place. Freesound is still one click away for effects.
A momentary provider hiccup no longer costs you a generation. The app gave up on the first failed status check — on work already created, already charged, and usually still rendering — and showed you the provider's own words for a fault you could do nothing about. It now tolerates a few before giving up.
You can get back to the brief and redraft. The only way in sat inside step one, which collapses and marks itself done the moment a storyboard exists — behind a tick that says there is nothing left to do there. It now shows on the step header, whether the step is open or not. It only reopens the brief; nothing is replaced until you press the button inside it.
A film whose lines are all recorded can still change its voice. Once nothing was left to record, the panel disappeared — and took the voice picker with it, so there was no way to pick a different voice and record again. The panel now stays, and offers to re-record instead of greying out. It says plainly that this charges again.
A clip you paid for can find its shot again. Clips whose shot link was lost mid-shoot finished, were charged, and were never attached — real footage sitting in your library with nothing pointing at it. Shots now carry that link themselves, and the ones already stranded have been reclaimed.
A refused character image says so, and says what was refused. The slot simply stayed empty, which against a one-at-a-time provider is indistinguishable from still waiting, and nothing said whether you had been charged. It now shows the reason and confirms the tokens came back. The old wording blamed your prompt; the check reads the picture as readily as the words and never says which, so it no longer guesses.
The assistant tab stops covering the labels beside it. It sits as a thin edge until you reach for it with the mouse or keyboard, then opens to full width.

0.35.0 — 2026-08-12

The frame you approve now has your actor in it

Stills are drawn from your cast photographs, text cards stop being filmed, and a generation that fails gives your tokens back.

Features:

Attach up to four reference images when generating an image, not one. The picture follows all of them — a subject, a style, a scene. Video generation has taken four for a while; images were limited by us, not by the model.
Seedance 2.5 is available in the video model picker. Takes up to 30 seconds in a single shot — twice anything else we offer — and generates sound alongside the picture rather than as a separate step.

Fixes:

The still you approve is now drawn from your cast photographs. It never was — the frame was generated from the words alone, so your actor's grey tee came back navy and a product's "neon green logo" wandered onto a shoe. Words steer a picture when the photographs are missing, and you were approving frames your actor had never been in. Redrawing a single frame carries the photographs too.
A text card is no longer filmed, and no longer charged for. A 90-second film that is mostly white text on black was quoted 190 tokens — most of that to have a video model produce black rectangles it is not even given the words for. Text belongs on the timeline, where it stays sharp and editable. Cards are now marked as cards, priced at nothing, and if we spot shots that look like cards we ask before changing anything.
A generation that fails gives your tokens back. Nine failures in ten days were charged and never refunded: whichever part of the app noticed the failure first settled it, and only one of those parts could pay you back. Tokens already lost have been returned.
Images generated from a Chinese prompt can be downloaded again. The filename is built from your prompt, and a single Chinese character was enough to make the download answer with an error instead of a file. It worked in English, which is why it lasted.

0.34.1 — 2026-08-11

Your words, if you wrote them in the right language already

Write your dialogue in Malay, pick a Malay voice, and the line reaches the camera exactly as you typed it.

Fixes:

If you write a line in the language you are filming in, it now reaches the camera exactly as you typed it. Translation was being asked to turn Malay into Malay, and instead of leaving it alone it tidied it up — "keyboard" came back as "papan kekunci", a casual phrasing came back formal, word order was rearranged. Three lines out of four in a real script, none of it asked for.

Changed:

How you write is treated as a choice, not a mistake. Casual stays casual, Malaysian Malay is not straightened into Indonesian, and a borrowed word you used on purpose stays the word you used. A script is approved in a particular voice and it should still be in that voice when it is filmed.
Choosing a voice in the language you already write in is now worth doing. There are Malay voices, and picking one means no translation happens at all. Picking an Indonesian voice for Malay text still translates it, which is what asking for Indonesian means.

0.34.0 — 2026-08-11

Films that speak your audience's language

Write in English, film in Indonesian — with mouths that match, lines that fit their shots, and products the size they really are.

Fixes:

A shot is bought long enough for the language it is actually spoken in. The spoken track was being generated from the English you typed and read out in an Indonesian voice, and the clip was sized from that English recording — but Indonesian takes about half again as long to say. Lines ran past the end of their own shot by up to nearly four seconds, which the edit cannot repair afterwards: the frames were never bought.
A product is filmed at its real size. In one ten-second cut a palm-sized speaker came back as large as the presenter's head, then correctly palm-sized in the next shot, then not a speaker at all but a strap over her shoulder. A product photographed against a plain background contains no clue about how big it is, and nothing in the shot was telling the model either. What you write on a product card now reaches the film, where before it was read only when making reference images.
Generating a video with several reference images could fail outright with a request-size error. Photos are now scaled and counted against a budget before they are sent.

Added:

Write your brief and your dialogue in English, give a character an Indonesian voice, and their shots are filmed in Indonesian. The line is translated before anything is generated and the translated line is what the model is given — so the mouth matches, because a single generation makes the picture and the speech together instead of laying one over the other.
You can read what your actor says. The spoken line sits under the English one on each shot card, so a language you do not speak is still something you can check. A translation nobody can read is a translation nobody can correct.
The whole film is translated in one pass rather than shot by shot, so a product named in shot two is named the same way in shot five. Translated separately they drift, and it is the kind of drift an audience notices without being able to say why. A line already translated is then reused rather than translated again — the same English can come back polite one run and street slang the next.

Changed:

If the dialogue cannot be translated, nothing is filmed and nothing is charged. It used to carry on and film the English under a prompt asking for Indonesian, without saying so. A film in the wrong language that looks finished is worse than one that was refused.
Mouth-matching is switched off, and an earlier release was wrong to announce it. The tool retimes a performance that is already talking, and our takes are deliberately closed-mouth, so there was nothing for it to retime — it returned each take unchanged and charged for the attempt. Measured, not assumed: 46 dB from the source, which is the noise floor of re-encoding. Filming the speech and the picture together is the answer, and it is the feature above.

0.31.0 — 2026-08-10

The mouths match

Each shot can be retimed to its own line, so a character stops miming while somebody else's voice plays.

Fixes:

Takes longer than ten seconds say so instead of failing at the provider. Shots are now generated long enough to hold their own dialogue, which means a wordy line can produce a take longer than matching can accept — the panel names how many rather than letting the attempt fail.

Added:

A shot can have its mouth matched to its line. Until now the voice was laid beside the picture and the two never agreed — the character moved their lips to nothing in particular while the track played over the top, which reads as bad dubbing whatever the language. Shoot, voice, then match: the take comes back with the mouth doing what the line says.
The original take is kept. Matching produces a second clip rather than replacing the first, so a result you dislike is dropped rather than re-shot — and a shot that cost four attempts to get right cannot be spent twice.

Changed:

Assembly uses the matched take when there is one and the original when there is not, so a film with no matching cuts exactly as it did before.

0.30.0 — 2026-08-10

The storyboard runs end to end

Frames you can approve, shots long enough to hold their own dialogue, and three steps that were quietly doing nothing.

Fixes:

Shots are bought long enough to say their line. A five-second shot with a seven-second line was generated at five seconds, and the missing two seconds could not be recovered afterwards — the frames were never bought. Length now follows the voice, so switching a film to another language re-cuts it instead of clipping it.
Shooting works. Every request was rejected before it reached the generator, and the failure looked exactly like having nothing approved yet, which is presumably how it went unnoticed.
Shots are generated from your cast references again. They were being sent without any — the step that keeps one face across a whole film was assembling its images and then not attaching them, which is the drift the cast set exists to prevent.
A character's clothes stay the colour you photographed. The written description outranked your own photo, so a beige top described as pastel-blue came back blue in two slots and beige in a third. The photograph now wins.

Added:

Storyboard frames are drawable. Every shot card offered a still and there was no way to make one — the panel priced them, the status machine expected them, and nothing anywhere could produce one. One token each, and they are the version you approve before any video is bought.
A reference set can be a product. The panel always said a character could be a person or a product; every prompt behind it asked for a portrait, so a keyboard came back as somebody's face. Person and product are now a choice on the card, and the object prompts never render a body.

Changed:

Clips are trimmed from the head rather than the tail, keeping the last moment a cut is built on — the gesture a match cut lands on, the whip pan into the next scene. A shot whose usable part is at the front can say so.
A shot with no clip yet holds its place with its frame instead of vanishing. Dropping it moved every later shot earlier and quietly pulled the film out of step with its own voice track.

0.29.3 — 2026-08-08

Rotation says so now

Turning something near square snapped it to square and told you nothing. It now shows the angle it landed on.

Added:

Rotate something to within a few degrees of straight and it clicks to straight — it always has. Now it says so, with the angle shown next to what you are turning. Before, the thing simply jumped and you were left deciding whether that was the tool helping or your hand slipping.

Changed:

Hold Shift while rotating and nothing snaps, which is unchanged — but nothing is shown either. If the label is not there, the angle is exactly where you put it.

0.29.2 — 2026-08-08

More than one voice

Chinese films had exactly one voice to choose from, so every character sounded like the same person. There are now fifty-six.

Added:

Languages with only a handful of voices now offer the multilingual ones alongside them. Chinese went from one voice to fifty-six, Indonesian from nine to sixty-four. It matters most where it was worst: with a single voice, every character in a film sounded like the same person.
They are listed after the voices belonging to the language, never in front of them, so whichever voice you had before is still the one you start with. English is untouched — it has over a thousand already and does not need the help.

0.29.1 — 2026-08-07

The tour could strand you on step two. Alignment guides drew one line where several edges were genuinely lined up. Both fixed, along with a sound setting that promised a language it could not deliver.

Fixes:

The tour could leave you stuck. If the step it was pointing at sat low on the page, the little card carrying Next and Skip was drawn past the bottom edge of the window — with the page dimmed behind it and no way to reach either button. It now stays on screen wherever the step is, and scrolls the thing it is describing into view first.
Alignment guides now show every edge that is really lined up, not just the one you happened to drag. Resize something centred and both sides arrive together, so both are marked. Drop a half-width element into the left half and it will tell you the outer edge is on the canvas edge and the inner one is on the centre — because both are true.
Resizing sometimes jumped to an alignment and drew nothing at all to say why. It was snapping correctly and simply not explaining itself, on more than half the snaps that moved anything.

Changed:

Generated sound is now off unless you ask for it. Turning it on hands the whole soundtrack to the model, voice included, and the model picks the language — so a clip could come back speaking English no matter what you wrote, and sound also forces a higher resolution, which costs more. Clips are quiet and cheaper by default now. For a voice in a language you choose, write dialogue in the storyboard and let Lucy speak it.

0.29.0 — 2026-08-07

The storyboard

Describe the film once. Lucy plans it shot by shot, and you see every shot as a still — cast, sets and voices settled — before anything is generated.

Fixes:

The motion box says what it is for. It describes how the camera and the subject move; a script pasted into it gets treated as directions and comes out as neither dialogue nor motion.
Pro mode clipped the storyboard and gave you no way to scroll to the rest of it.

Added:

Tell Lucy what the film is and she drafts the whole of it: the shots, what happens in each one, who is in it and what they say. You start by arguing with a plan instead of staring at an empty timeline.
Every shot appears as a still before it is a video. This is the point of the whole thing — a shot you do not like costs you a sentence to fix at this stage and a regeneration after it.
Cast and sets. A character or a location gets its reference image once, and every shot they appear in uses it, so the same person still looks like the same person by shot nine.
Nothing is shot until you have approved the cast. Changing a reference after you have approved it asks you again rather than quietly going ahead with the face you already rejected.
The dialogue gets spoken out loud. Lines written in the plan become real speech in the language the script is written in, and land on the timeline against the shot they belong to.
A voice for each character, chosen from the full catalogue — not one narrator reading every part.
The storyboard is in pro mode as well as easy mode. Same storyboard and the same wiring underneath, so a film started in one loses nothing when you open it in the other.
The tour covers the storyboard now, in whichever mode you are working in and whichever language you read.

0.28.0 — 2026-08-06

Sign in with Google

Continue with Google on the sign-in and sign-up pages. If you already have a Lucy account on the same address, it is the same account — not a second one.

Added:

Sign in with Google, on both the sign-in and sign-up pages. The first time creates your account and every time after signs you in, which is why the button says continue rather than either.
If you already signed up with an email and password, using Google on that same address signs you into the account you already have. Your tokens, projects and history stay where they are — you do not end up with two accounts holding half your work each.

0.27.0 — 2026-08-06

Describe the film. Lucy plans it.

Write a brief and Lucy drafts a shot list — camera, dialogue, captions, transitions — laid out as a storyboard you can read, edit, and send to a client before a single second of video is generated. Approve it, and she shoots it.

Improvements:

Generations that run for minutes now survive you closing the tab, whether they are reference images or shots. They keep going, they charge or refund correctly, and they find their way back to the right slot when you return.

Fixes:

A generated image or video could not be saved to your computer. The result had an "Add to media" button and nothing else, and because the file is served from elsewhere, the browser's own download option did nothing either. There is now a download button, and the file keeps the name it has in your library.

Added:

Write a brief in easy mode and Lucy drafts the whole film as a shot list — how each shot is framed and moves, who is in it, what is said, what appears on screen, and how it cuts to the next. It arrives in about fifteen seconds and every field can be edited in place.
The shot list is laid out as a storyboard, not a settings panel. It is meant to be read by someone who has never opened Lucy — which is the point, because this is the thing you send a client before anyone spends money filming.
Reference sets keep a character looking like the same character. Upload one clear front-on photo and Lucy builds the other three angles from it. The four slots are not interchangeable and the panel says what each is for — the front-on shot anchors the identity, the three-quarter angle teaches the face turning away, and the wardrobe close-up is what stops a jacket changing colour between cuts. It works for a product as well as a person.
Shoot the shots you have approved and they assemble themselves onto the timeline in order, trimmed to the length you approved, as a single step you can undo. Where a shot said how it ends — a fade, a dissolve, a whip pan — that becomes the transition. Where it said something a transition cannot express, like a match cut, it stays a cut and keeps your note.
Every price is shown before you commit, never after you are charged. A still costs a token and a whole film costs about sixty, which is the entire reason the storyboard comes first: a change costs one token at this stage and sixty after it.

0.26.1 — 2026-08-06

Telegram fixes

The bot had stopped replying, and it refused images sent as files — which is how Telegram sends them when you drag one in from your desktop.

Fixes:

The Telegram bot stopped replying. Messages sent to it went nowhere for several hours — they were being delivered to the wrong place. Anything you sent during that window was queued and has since been answered.
Sending the bot an image as a file was refused. Telegram sends images as files rather than photos when you drag them in from a desktop, so this affected the most common way of getting a picture to Lucy. Files now work, and they are the better input — they skip Telegram's compression, which matters when the image is a reference for a character who has to look the same in every shot.

0.26.0 — 2026-08-06

Record, sign in, and stop losing work

You can now record a voiceover, your camera or your screen without leaving Lucy. Generations keep running when you navigate away. There is a way back in from a forgotten password, a way in with Telegram, and every token you spend is now accounted for.

Improvements:

Reverse playback holds up on longer clips, where a limit tuned for short ones caused it to stall.
Generated music and sound effects now reach your media library like everything else.
Several corners of the editor were still in English regardless of the language you had chosen.

Fixes:

Finished videos could be unreachable. A generation that completed while you were elsewhere was charged for, rendered, and then offered nowhere in the app. Those are now available to add, including ones from before this update.
Generated images and videos are added to your media library automatically instead of waiting for a click.
Screen recordings came out unwatchable. They were being encoded into a format that tops out well below the size of a modern display.
"Invalid generation request" told you nothing. When a generation is refused it now names what was wrong — the model, the length, the reference, or your balance.
The account page fell apart once you had used it for a while. A long token history pushed security, Telegram and everything else off the bottom of the screen. It is now two columns, and the history stays in its own panel.
The account page ignored your theme and was always dark.

Added:

Record a voiceover, your camera, or your screen. Voiceover plays the timeline while you narrate and drops the take exactly where you started, so it lands against the picture rather than at the beginning. Pause and resume mid-take.
Generations keep running when you navigate away. Starting a video and then leaving the page used to lose it — the render finished and you never got it. The work now continues on our side and appears when it is done, even if you close the tab.
A status bar along the bottom of the editor shows what is running, with the project frame rate and how well playback is keeping up. Click it to see everything in progress.
Sign in with Telegram, and reset a forgotten password. Neither existed before; a lost password meant a lost account.
Up to four reference images per video, so a character can hold together across shots. The four slots are not interchangeable and the panel now says what each is for — a frontal portrait, a three-quarter angle, and a wardrobe close-up do far more than four near-identical selfies.
Click your token balance to see where the tokens went. Every charge and refund, what it was for, and when.

0.25.0 — 2026-08-04

Fixes, mostly

A round of repairs to the things that shipped last week — adding a GIF no longer hangs, long GIFs play all the way through, generated items have names you can read, and the Telegram bot finally listens when you ask for a shape.

Improvements:

Reverse playback is smoother at 4K and 60fps, where a limit meant to protect the preview was never taking effect. It is not perfect — footage with long gaps between keyframes can still stutter in reverse.
The GIF grid loads faster and uses far less data, and searches you have already run come back instantly.

Fixes:

Adding a GIF could spin forever. Picking a sticker sometimes never finished, with no error and no way out. It now adds immediately, and a slow image can no longer hang the button.
Long GIFs played only partway through. A sticker with more frames than we budgeted for stopped early and looped from there, silently — so what you saw in the grid was not what you got on the timeline. The whole animation plays now.
Generated images and videos showed a string of digits instead of a name. Each one is now named after the prompt that made it, in the media folder and on the timeline, including the ones you generated before this update.
Generated images could not be opened at full size. Click one to see it properly before you commit it to your project.
The Telegram bot ignored the shape you asked for. Writing 16:9 or 9:16 in your message had no effect and everything came back square. It now reads the ratio from what you type, and still defaults to square when you do not ask.
Regenerate could fail without saying anything. When it could not run, the button simply did nothing; now it tells you why.
In light theme, a white bar sat across the top of the page and Sign out was invisible against the background.

0.24.0 — 2026-08-02

Stickers that move

GIPHY brings millions of stickers and GIFs to the Stickers panel — and, for the first time, stickers that animate in the preview and in the exported video.

Features:

GIPHY. Search millions of stickers and GIFs from the Stickers panel and drop them straight onto the timeline. Transparent stickers come first, because they composite over your footage cleanly.
Animated stickers. Until now a sticker was a still image. Now a GIF plays — in the preview and in the exported video, showing the same frame at the same moment in both.

Improvements:

A sticker you place today keeps working. The link is rebuilt from the sticker itself, so reopening an old project needs no trip to GIPHY.

0.23.0 — 2026-08-01

Make a face sing

Avatar turns a photo and an audio clip into a talking, singing face — and vocal removal splits any song into an instrumental and a vocal track, free, on your own machine.

Features:

Avatar. Give it a face — a photo or a video clip — and up to 30 seconds of audio, and it generates the face singing or speaking it. The cost is shown per second before you commit, so there are no surprises.
Split vocals. Separate any track in your project into an instrumental and a vocal, right on your own machine — free, no tokens, and nothing is uploaded. Use the instrumental as a karaoke backing with one click. It removes most of the vocals; quiet traces can remain on dense mixes.

Improvements:

AI sound effects are ready behind the scenes and will appear once enabled.

0.22.0 — 2026-08-01

Pick your model, watch it work

Choose which AI model generates your video, get longer and more flexible clips with Kling 3.0, and actually see how far along a generation is.

Features:

Choose your video model. Kling 3.0 is now the default and brings genuinely more range: any length from 3 to 15 seconds instead of only 5 or 10, and up to seven reference images instead of one. Older models stay available, and the controls change to match whichever you pick — you can't pick a setting the model won't accept.

Improvements:

Generation now shows real progress instead of an endless spinner. You see the actual percentage as it goes, and whether it's queued, generating or downloading — so a job waiting for a slot no longer looks broken.
Generated video is 1080p when it includes sound, and video length now offers exactly the durations the chosen model supports.

0.21.0 — 2026-08-01

Generated video, with sound

AI video now comes back with audio instead of silence, and our privacy and terms pages have been rewritten to describe honestly what happens to your data.

Features:

Generated video now has sound. Clips come back with audio matched to the scene rather than silent, rendered at 1080p. It costs no more than before.

Improvements:

Clearer, more honest privacy and terms. The old pages said your content never leaves your device, which was inherited from the project Lucy grew out of and hasn't been true since AI features arrived. They now spell out exactly what stays on your machine — your projects, media, exports, captions and background removal — and precisely which service sees what when you use a feature that needs one.

Fixes:

Video length now only offers durations that actually work. Asking for other lengths used to be accepted and then fail.

0.20.0 — 2026-08-01

Music, properly licensed

A music library arrives — thousands of tracks you can actually use — along with a fix for sound effects that were being offered as cleared when they weren't.

Features:

A Music tab with thousands of royalty-free tracks. Filter for instrumentals, preview before adding, and use Fit to project to trim a track to your edit with a gentle fade-out — or loop it if it's shorter.
Every track shows its licence plainly. Tracks that need the artist credited say so before you add them, and a new Credits view lists everything in your project that needs crediting — with the text ready to paste into your post.

Improvements:

A link to Pixabay's music library, whose tracks are free to use without crediting anyone. Download there and import via the Media panel.

Fixes:

Sound effects marked as cleared for commercial use weren't always filtered correctly, so tracks restricted to non-commercial use could appear alongside the rest. That's fixed: non-commercial and no-derivatives licences are now excluded properly, and anything we can't verify is excluded by default.

0.19.2 — 2026-08-01

Portrait, actually

Asking for a 9:16 image now gives you a 9:16 image. It has been quietly returning squares.

Fixes:

Choosing an aspect ratio for AI image generation had no effect — every image came back square no matter what you asked for. Portrait now returns portrait and landscape returns landscape, at the size you chose.

0.19.1 — 2026-07-31

Where did that clip go?

Adding a smart clip looked like it did nothing — it was landing at the end of a long timeline, far off screen.

Fixes:

Adding a suggested clip appeared to do nothing on longer projects. The clip was in fact being added correctly — at the end of the main track, which on a long timeline is well past what you can see. Lucy now moves the playhead to the new clip and selects it, and the message tells you where it landed.
A suggestion with no length could produce an invisible, zero-length clip. Those are skipped now, and if nothing could be added you get an error instead of a success message.

0.19.0 — 2026-07-31

Sized for the feed

Export presets for YouTube, TikTok, Instagram and X — including a choice of how your frame fills a vertical video.

Features:

Export presets. Pick YouTube, Shorts, TikTok, Instagram Reels, a square or portrait Instagram post, or X, and the resolution, frame rate and encoding are set for you. Custom is still there when you want the manual controls.
Landscape edit, vertical export? Choose how the frame is filled: keep everything visible with bars, crop to fill the frame, or fit the video over a blurred copy of itself. A preview shows the result before you export, and your project is never altered — export the same edit for as many platforms as you like.

Improvements:

If your video runs longer than a platform allows, Lucy tells you the actual numbers before you export instead of letting you find out after uploading.

0.18.0 — 2026-07-31

Ready to post

Finish an export and Lucy hands you the file, a caption ready to paste, and a link straight to wherever you're posting.

Features:

Share your video after exporting. Write a caption, copy it with one tap, and open the upload page for YouTube, TikTok, Instagram, or X. Your file stays on your machine — Lucy hands you everything you need to post, you stay in control of where it goes.
Videos shared from a free account carry a short "Made with Lucy" line in the caption. It is always shown before you copy, never hidden, and Creator and Studio subscribers can switch it off.

Improvements:

Links shared from Lucy now show a proper preview card.

0.17.0 — 2026-07-31

Every project, at a glance

Projects now show a real thumbnail — pick any frame from your video or upload your own image — and the product is now simply called Lucy.

Features:

Project thumbnails. Your projects list shows a real frame instead of a placeholder. Open Set thumbnail to scrub through the video and pick the exact frame you want, or upload your own image. Projects without one get a thumbnail automatically, taken from the first frame that actually has something in it.
Thumbnails are rendered through the same pipeline as your exports, so the frame you pick is the frame you get — and like all your footage, the images stay on your device.

Improvements:

Tablets held upright now get a quiet, dismissible hint suggesting landscape for more room to work.
The product is now called Lucy, everywhere — in the app, in emails, and in your browser tab.

0.16.1 — 2026-07-31

Room to work

The editor now asks phone users to switch to a larger screen — and tablets are welcome, where before they were turned away too.

Improvements:

Tablets can now open the editor. The old size check turned away anything under 1024px wide, which included every iPad in portrait; the limit is now 768px, so tablets work and only phone-sized screens are asked to switch to a larger display.
The message is clearer, appears in your language, and the editor no longer loads in the background on a screen that cannot show it. Resize a window or rotate a tablet across the threshold and the editor appears or steps aside immediately, with no reload.

0.16.0 — 2026-07-31

说你的语言

Lucy now speaks Simplified and Traditional Chinese, with a language switcher in the header, the editor, and your account settings.

Features:

Simplified and Traditional Chinese. The landing page, sign-in, pricing, account, projects, and the editor's main controls are all translated, with terminology chosen for each region rather than converted character by character — 專案 and 影片 in Traditional, 项目 and 视频 in Simplified.
A language switcher sits in the site header, the editor menu, and your account page. Your choice is remembered, and first-time visitors get their browser's language automatically.

Improvements:

A new Lucy mark replaces the placeholder OpenCut logo across the site, browser tabs, and home-screen icons.

0.15.1 — 2026-07-31

Sound on the first beat

Playback starts with audio immediately instead of running silent for a couple of seconds after every edit.

Fixes:

Pressing play could start silent, with sound only arriving after a scrub — and it happened again after every pause and resume. Any timeline change was throwing away decoded audio, so the next play had to re-decode whole clips before making a sound. Audio now starts within a frame of pressing play (about 0.04 seconds, down from two seconds), and clips with fades stream with their volume curve applied live instead of waiting on a full decode.

0.15.0 — 2026-07-31

正體中文

Chinese captions can now come out in Traditional characters — by default — and existing captions convert in one click.

Features:

Traditional Chinese captions. Speech recognition tends to produce Simplified characters; Lucy now converts them for you. A Chinese script preference on caption generation defaults to Traditional and remembers your choice.
Convert captions you already have: right-click a caption and choose Convert to Traditional (or Simplified) — every caption converts at once, as a single undoable step. Mixed text is safe: Latin words, numbers, and already-Traditional characters pass through untouched.

0.14.2 — 2026-07-31

Featherweight timeline

Playback smoothness, part two: the timeline no longer competes with your video for frame time, and oversized sources decode leaner.

Improvements:

The timeline used to redraw all its clips, ticks, and waveforms on every single frame of playback — on caption-heavy projects that alone consumed the whole frame budget. It now updates only the playhead during playback (roughly 90× cheaper), so busy projects play smoothly.
Sources bigger than your project — like a 4K clip in a 1080p edit — now decode at project resolution during preview, cutting memory use 4×. Exports still use the full original quality.

0.14.1 — 2026-07-31

Full speed backwards

Ping-pong playback is now perfectly smooth — reverse segments play at full frame rate instead of stuttering.

Improvements:

Ping-pong preview now plays reverse segments at full frame rate. Video decodes forward by nature, so playing backward used to stutter around 10fps — clips now decode ahead into a small buffer and play back from memory, making reverse indistinguishable from forward.
Ping-pong exports render several times faster, and previews with a blurred background do half the decoding work they used to.

0.14.0 — 2026-07-31

Round and round

Stretch a video past its end and it loops — or ping-pongs, playing forward then backward for a truly seamless endless loop.

Features:

Video looping. Set a clip to Loop and drag it as long as you need — the video repeats, audio and all, with small markers showing each loop point on the timeline.
Ping-pong mode plays the clip forward, then backward, then forward again — no jump cut at the seam, ever. Perfect for patterned background loops and karaoke productions. Source audio is muted in ping-pong (your music track plays on unaffected).

Improvements:

Speed and retime settings compose with looping, so a slowed or sped-up clip loops exactly as it plays.

0.13.0 — 2026-07-30

Ease in, ease out

Audio gets proper fade in and fade out with smooth equal-power curves, and the spacebar always comes back after editing text.

Features:

Audio fades. Give any audio clip — or the sound of a video clip — a fade in and fade out. The curve is equal-power, so fades sound smooth instead of dropping off a cliff, they stack cleanly with your volume settings and keyframes, and exports match the preview exactly. Small ramps at the clip edges show fades at a glance.

Fixes:

After editing text, clicking back into the timeline now always returns keyboard control — spacebar plays and pauses again without having to click an asset first. Clicking away mid-edit also commits your text instead of quietly discarding it.

0.12.1 — 2026-07-30

Steady frames

Two fixes from real editing sessions: no more black flashes while scrubbing, and cleaner undo history for color edits.

Fixes:

Scrubbing the timeline could momentarily flash black over video while frames decoded. The preview now always holds the last frame until the next one is ready — and exports can no longer emit a black frame in those spots either.
Changing a color sometimes added a duplicate entry to your undo history, making undo feel like it did nothing. Every color edit is now exactly one undo step, and picking the same color again adds none.

0.12.0 — 2026-07-30

Captions, together

Select every caption at once and style them in one move, text editing that never drops a keystroke, and Telegram learns to work from your photos.

Features:

Style all your captions at once. Right-click any caption and Select all captions, then set color, position, effects, and animations for the whole set in one go — mixed values are labeled, and a single undo reverts the whole batch.
Send @lucyaidobot a photo. A photo with a caption generates an image from your picture; caption it /video and the photo becomes the opening frame of a generated clip. Results come back in chat and land in your editor inbox.
The Telegram mini app can attach a reference too — upload a picture or take one with your camera and generate from it on the go.

Fixes:

Text editing no longer drops your changes. Edits to a text layer's content could vanish mid-typing when the editor refreshed underneath you — typed text now always sticks, in the panel and directly on the canvas.

0.11.0 — 2026-07-30

Start from a picture

Generation now takes a reference image — hand Lucy a picture and ask for an image in its style or a video that starts from it, in both Easy and Professional modes.

Features:

Reference images for image generation. Attach any picture — upload it, drag it in, paste it, or pick one from your library — and describe what you want: a watercolor version, the same character somewhere new, your product in a different scene. Seedream generates from your reference for the same 1 token.
Reference images for video. The first-frame picker now accepts uploads and drag-drop too, so any image on your machine can become the opening frame of a generated clip — not just ones already in your library.
Easy mode gets a paperclip. Attach a reference right in the prompt dock and the whole still-to-motion flow builds on your picture.

Improvements:

Oversized references are downscaled automatically before upload, so large photos just work.

0.10.0 — 2026-07-29

Motion and type

Text gets drop shadows and outlines with one-click presets, and every layer can now animate in, out, and in place — with effects that stack cleanly on top of your own keyframes.

Features:

Text drop shadow and stroke. Give any text layer a soft or hard shadow (color, offset, blur, opacity) and a clean outline (color, width) — with one-click presets like Soft shadow, Hard shadow, and Outline. Every value is keyframeable and renders identically in exports.
Element animations. Pick an In, Out, and Loop animation for any text, image, video, or sticker layer: fade, four slide directions, pop, and a typewriter reveal for text, plus pulse and drift loops. Durations and speed are adjustable after the fact, and removing one is a single click.
Animations compose with your keyframes instead of replacing them — a clip can slide in while your own scale and position animation keeps running underneath. If a clip already has a transition on an edge, the transition wins and the preset stays out of the way.

Improvements:

Smoother slider controls across the properties panel — dragging feels more precise and commits as a single undo step.

0.9.0 — 2026-07-29

Lucy in your pocket

Lucy now lives in Telegram too: generate images and clips from chat with @lucyaidobot, use the Open Lucy mini app without leaving Telegram, and pull everything you made on the go straight into your projects.

Features:

Meet @lucyaidobot. Link Telegram from your account page, then generate from any chat: /image makes a still, /video makes a clip, /balance shows your tokens, /history lists your recent creations. Finished results are delivered right back into the chat.
The Open Lucy mini app. Tap the bot's menu button to open a pocket-sized Lucy inside Telegram — check your balance, browse your generation history with previews, and start new image or video generations without leaving the app.
The editor inbox. Anything you generate outside the editor is waiting for you when you next open a project — one click adds it to your media library with its prompt attached, into whichever project you're working on.

Improvements:

If a generation fails on our side, your tokens are refunded automatically — in the editor, in the bot, everywhere.

0.8.0 — 2026-07-28

Image tools

Two new tools for image layers — background removal that runs entirely on your device, and AI object removal — plus fairer, per-second pricing for video generation.

Features:

Remove background. One click cuts the subject out of any image and gives you a transparent PNG. It runs entirely on your device: free, private, and unlimited. The first use downloads the model; after that it takes seconds.
Remove object. Describe what should disappear — a stray sign, a photobomber, floating leaves — and Lucy edits it out of the image while keeping everything else in place. 1 token per edit.

Improvements:

Video generation is now priced by length: 2 tokens per second, so a 5-second clip costs 10 tokens and a 3-second one costs 6. The cost is shown before you generate.

0.7.0 — 2026-07-27

Smart clips & stock

Lucy now helps you find the good parts. On-device AI picks the strongest moments out of long footage and trims the silences, and a built-in stock browser brings free photos and videos into your project.

Features:

Smart clips. Point Lucy at a long recording and it finds the strongest moments and suggests ready-to-use clips. Analysis runs entirely on your device — nothing is uploaded and it costs no tokens.
Silence trimming. One click removes the dead air from a talking clip while keeping natural pauses intact.
Stock library. Search free photos and videos from Pixabay and photos from Unsplash right in the editor, preview them, and add them to your project with proper credits shown.

Fixes:

Adding your first clip no longer silently changes an explicitly chosen canvas size. If you haven't picked one, the canvas still matches your first clip — but now it tells you and offers a one-click Revert.

0.6.0 — 2026-07-26

The cutting room

This release is for the editors: real transitions between clips, color grading with adjustment layers, precise transform controls with keyboard shortcuts, and your own fonts.

Features:

Transitions. Seven of them — crossfade, fade through black, zoom, slide left, slide right, wipe left, and wipe right. Drop one between any two adjacent clips, adjust its duration, and it renders in the preview and the export alike. Works on the main track and on overlay video tracks.
Color grading. A new color adjust effect gives every video and image layer brightness, contrast, saturation, and tint controls — keyframeable like everything else.
Adjustment layers. Put a grade on its own layer and it applies to everything beneath it. Stack them, trim them, fade them.
Transform keyboard shortcuts. Nudge position with Alt+arrows, scale and rotate from the keyboard, and reset a transform in one stroke. All of them are rebindable in the shortcuts editor.
Custom fonts. Upload your own TTF, OTF, or WOFF files and they appear in the font picker next to Google and system fonts — stored with the project and included in exports.

Improvements:

Smoother preview playback when the same clip appears several times on the timeline.

0.5.0 — 2026-07-25

Easy mode

A second way to edit. Easy mode turns Lucy into a guided storyboard: type what you want to see, get a still, turn it into a clip, and keep going — no timeline knowledge required.

Features:

Easy mode. A storyboard view where every scene is a card. Write a prompt, Lucy makes the still, then turns it into a clip. Switch between Easy and Pro at any time from the header — it's the same project underneath.
The prompt dock. One input at the bottom of the screen drives everything: describe your next scene and it's added to the story. Costs are shown up front before you generate.
Per-card actions. Regenerate a scene you don't like, or extend it with a follow-up clip, straight from the card.
One-click captions in Easy mode. Lucy transcribes your video on your device and drops styled captions onto the story — no upload, no tokens.
A short first-run tour walks new editors through the storyboard the first time they open it.

0.4.0 — 2026-07-24

Meet Lucy

The editor you know, reborn as Lucy — with AI image and video generation built in, free accounts, and a token balance that only ever pays for generation.

Features:

Generate still images from a text prompt without leaving the editor. Describe the shot, pick an aspect ratio, and the result lands straight in your project's media library.
Bring a still to life. Any generated image can be turned into a short video clip with a motion prompt — great for establishing shots, b-roll, and moodpieces.
Accounts and tokens. Sign up, verify your email, and get 100 free tokens. A still costs 1 token; editing itself never costs anything and never will.
A new account page shows your token balance and full generation history, and lets you delete your account — along with all of its data — at any time.

Improvements:

Generation runs in the background while you keep editing. Progress is visible on the asset card, and finished results are saved into the project like any other media.

0.3.0 — 2026-04-15

Masks, animation & more

This release adds masks, a graph editor for keyframe curves, volume and speed controls, preview zoom, canvas backgrounds, stickers, and a lot more.

Features:

Added a brand page with downloadable brand assets. Assets are also accessible directly from the header logo context menu.
Scale now has separate width and height controls. You can stretch or squash elements independently on each axis.
Volume control for audio and video elements.
Speed control for video and audio clips. Includes a maintain pitch option so audio doesn't chipmunk or deepen when you change the rate.
Masks. Hide or reveal parts of any video or image layer. Choose from split, rectangle, ellipse, star, heart, diamond, and cinematic bars. Adjust position, size, rotation, feather, and stroke in the properties panel, or drag the handles in the preview.
Canvas backgrounds are back. Set a blur, solid color, or gradient as the canvas background from the new Background tab in the settings panel.
Custom canvas size. Instead of being locked to a handful of presets, you can now enter any width and height you want.
Preview zoom and panning. Zoom into the canvas to work on fine details, and pan around to navigate when zoomed in.
Stickers panel now has content. Browse and add country flags and shapes.
You can now import transcript files to generate captions instead of running auto-transcription.
Graph editor for keyframe curves. Shape the easing between keyframes by dragging bezier handles.
Multiple selected timeline clips move and resize together.
Position X and Y can now be animated independently.
Keyframes can be copied and pasted between elements.
Timeline clips can expand to show a lane for each animated property. Right-click a clip -> "Expand keyframes" to see all properties in their own rows.

Improvements:

Added a changelog link to the footer.
Playback performance is dramatically better. The editor used to slow to a crawl during playback, and audio would stutter when interacting with elements mid-play. Both are fixed.
More vertical space between timeline tracks. Selection outlines on adjacent tracks no longer overlap each other.
Holding Shift or Ctrl while drag-selecting in the timeline now adds to your existing selection instead of replacing it. Matches how Shift/Ctrl+click already worked.
Timeline track rows now highlight when they contain a selected element, making it easier to see which tracks are active at a glance.
System fonts (Arial, Helvetica, Times New Roman, and others) now appear in the font picker alongside Google Fonts.
Your projects can disappear if your disk runs low on space. Lucy now asks the browser to protect them.
The properties panel has been redesigned with a tabbed layout.
Audio waveforms have gotten an upgrade. The old ones were inaccurate, prone to crashing, and not very useful. The new ones are RMS-based, properly scaled, and only render what's visible so they don't slow things down.
The timeline now opens scrolled to the bottom, so your main video track is visible right away instead of being hidden above the fold.
The settings panel has been redesigned with a cleaner layout and better organization.
You can now select multiple assets in the assets panel. Useful for deleting several at once.
If your browser doesn't support GPU-accelerated rendering, the editor now shows a notice explaining why and suggests switching to another browser.

Fixes:

Sponsor logos now correctly invert in dark mode.
Media asset cards in the grid view were larger than their thumbnails, leaving empty space around each item. Cards now size to their content.
Project thumbnails now generate correctly even when the timeline is empty. If you set a background color or gradient with no clips, the thumbnail reflects that instead of staying blank.
Clicking a selected element no longer keeps the rest of the selection.
Track labels on the left side of the timeline were misaligned with the tracks on the right.
Clicking in the empty space below timeline tracks now seeks the playhead.
Timeline vertical scroll now works anywhere in the panel, not just over the track labels. The header no longer scrolls with the tracks.
MP4 exports with audio failed on Firefox. The export now completes successfully regardless of browser.
Snap guides in the preview panel didn't show when moving a full-canvas element to the center of the frame.
The rotation handle could float above the editor UI when an element was near the top of the canvas. It now clips to the preview area.
Images with a blur effect at 0% intensity would disappear completely when a blurred background was also active. Now 0% blur correctly renders the image unchanged.
Section titles in the properties panel showed the wrong text color when the section wasn't collapsible. Non-collapsible titles now use the correct muted style.
Blur looked pixelated at higher intensities instead of producing a real blur. Both the canvas background blur and the blur effect are now fixed.
Auto-generated captions were sometimes inaccurate. The underlying issue has been fixed.
Text element handles no longer shift position as you type.
The rotation handle drifted further and further from the element as the canvas size got smaller. On a 100x100 project it was almost off-screen. The gap is now a fixed size regardless of canvas dimensions or zoom.

Technical details:

Ripple editing now has a proper algorithmic foundation. The old inline approach was replaced with a diff-based system in src/lib/ripple/: track snapshots before and after an operation are compared as interval sets, vacated intervals are computed (accounting for elements moved to other tracks), and the result is a list of explicit RippleAdjustment values applied in a separate pass. Composable, testable, and independent of the command that triggered it.
Time is now represented as MediaTime, an integer tick count at 120,000 ticks per second, defined in rust/crates/time. Previously, floating-point seconds were used throughout, which accumulated rounding errors and made frame alignment imprecise. 120,000 was chosen because it divides evenly by every standard frame rate denominator, including drop-frame rates: 23.976 maps to 5,005 ticks/frame, 29.97 to 4,004, 30 to 4,000. FrameRate is a rational {numerator, denominator} type rather than a float. All time arithmetic, frame snapping, and timecode formatting are implemented in Rust and exposed via WASM.
Track ordering is now enforced at the type level. SceneTracks replaces a flat track array with three explicit fields: overlay, main (a singleton VideoTrack), and audio. The type makes it structurally impossible to mix track kinds or misplace a track.
The WebGL renderer has been replaced with a Rust/wgpu compositor compiled to WASM. The rust/ crate tree covers compositing (rust/crates/compositor), effects (rust/crates/effects), masks with JFA-based feathering (rust/crates/masks), GPU context (rust/crates/gpu), and time (rust/crates/time). WASM bindings live in rust/wasm/ and are consumed from TypeScript through src/lib/wasm/. The old GLSL shaders and WebGL utilities are gone.
src/types/ and src/constants/ have been dissolved. All types and constants now live alongside the code that owns them inside src/lib/. The packages/env and packages/ui internal packages were also folded into the web app, with icons at src/components/icons/ and env at src/lib/env/.
The web app can now deploy to Cloudflare Workers via OpenNext. wrangler.jsonc and apps/web/open-next.config.ts are included for anyone self-hosting.
apps/desktop/ is a new GPUI + Rust app. It is early: there is a main.rs and a working build setup, but no features yet.
16 new storage migration steps were added (v9 through v25), covering the schema changes accumulated during this release cycle.

0.2.0 — 2026-03-01

Motion & effects

This release adds the foundation for motion and effects, alongside many improvements & fixes.

Features:

Keyframe animation system for animating element properties over time.
Entirely new efects system. Effects can be applied per-clip or added as standalone elements on the timeline (with our first effect: Blur!)
Added a changelog page to the site.
Ripple editing mode.

Improvements:

New text elements no longer have a black background by default.
Selection outline in the preview is now dashed.
Clips sitting next to each other on the timeline now have a small gap between them so selected clips are always clearly visible.
Press Escape to deselect the current element.

Fixes:

Fixed an issue where click-and-drag selection on one timeline track could accidentally select items from the track below.
Audio could flicker or cut out at the start of playback when certain elements were selected. The scheduler now recovers from timing slips without losing audio.
Dragging a clip onto the timeline and pressing Ctrl+Z now removes both the clip and the track it created in one step, instead of requiring two undos.
A few values in the properties panel were previously showing incorrect values. This has been fixed.
Timeline trim handles now sit outside the clip instead of inside them, so the clip's visible area accurately represents where it starts and ends.
Clips being dragged on the timeline now render on top of other clips instead of underneath them.
Undoing a trim-from-start operation now correctly restores the clip to its original position and length, instead of stretching it to the right.
Resizing a clip on the timeline would deselect it when the mouse was released. The clip now stays selected after a resize.
Resizing a clip could extend it past another clip on the same track, causing an overlap. The resize now stops at the next clip's edge.
Dragging a clip to a different track would deselect it on drop. The clip stays selected after moving.
Finishing a resize or drag with the mouse over empty track space would deselect the clip.

0.1.0 — 2026-02-23

Editor foundation

This first release focuses on making editing faster, clearer, and more reliable for day-to-day use.

Features:

Rebuilt the properties panel from scratch. Number inputs now support math expressions and click-drag scrubbing instead of just typing values in.
The properties panel went from a flat list of basic text styles to organized sections: transform, blending, typography, spacing, and background controls. Text elements finally have position, scale, and rotation.
New color picker with eyedropper, opacity control, and multiple color formats.
Bookmarks now support notes, color labels, and optional duration to plan scenes more clearly.
Move, scale, and rotate elements directly in the preview panel.
Blend modes (Multiply, Screen, Overlay, etc). Only available on text elements for now.
Paste images, videos, or audio from your clipboard straight into the editor.
Hold Shift while moving or resizing to temporarily turn off snapping.

Improvements:

Expanded font selection from 7 to over 1,000.
Fonts load much faster.
All asset panel tabs now share a consistent layout.
Text rendering properly supports multiple lines, line height, and letter spacing.
Better contrast in the dark theme.

Fixes:

Fixed undo/redo for drag-and-drop so dropped items are reliably undoable.