Edit photos by describing the change, on your own GPU. A free, open-source alternative to Nano Banana and ChatGPT image editing: photos stay on your computer, no daily limit. One installer for Windows, Mac and Linux sets up the best open model (Qwen-Image 2.1). MCP server for AI agents.
Marketing
Launch material for Kokon. Everything visual here was made by Kokon itself or by its own UI code.
| Folder | What |
|---|---|
showcase/ |
The before/after pairs and the hero image, generated with Kokon (prompts, seeds and times in PROMPTS.md). Web-sized JPEGs; the full-size PNGs stay in %USERPROFILE%\Pictures\Kokon\marketing. |
video/ |
The launch video, made with Remotion from the app's own React components. |
release-notes-0.9.0.md |
Body of the v0.9.0 GitHub release (already in the draft). |
../site/ |
The one-page website (static HTML, no build step). |
The video
video/ renders the real Sidebar, Stage, Composer and timeline from ../src. Webpack swaps
the app's store, service API, native bridge and Tauri modules for mocks (video/src/mocks), so
each frame's UI comes from a state computed for that frame. Change the app's UI and the video
follows on the next render. Two parts are copies with the same markup and classes, because
their state lives inside the component: the paint-area editor (app/InpaintOverlay.tsx) and
the before/after slider (mocks/Compare.tsx). The job card's motion (percentage, Prompt+ text,
cocoon, hatching) runs on timers in the app; in the video mocks/jobMotion.ts and
mocks/Cocoon.tsx drive the same code (src/components/jobMotion.ts, drawCocoon, paintHatch)
from the frame, through <AppWindow card={…}>, since Remotion renders frames out of order. CSS
loops follow the frame through --vt (app/theme.css).
cd marketing/video
npm install
npm run studio # preview and scrub in the browser
npm run render # out/kokon-launch.mp4, 1920×1080, ~85 s, about 5 minutes
npx remotion render KokonLaunch out/kokon-launch-0.10.mp4 # same, to a new file (keep old renders)
Scenes (src/scenes): Intro (the mark assembles) → Hook (wall of before/afters) → Create (a few
words, Prompt+ writes the prompt, the image hatches) → Region (paint, remove, add) → Versions →
Montage (five edits, full frame) → Agent (MCP) → Hardware (Windows, macOS, Linux) → Outro. Each scene is also its own composition, so one can be rendered alone, for example
npx remotion render Montage out/montage.mp4 (the voice and music play only in KokonLaunch).
"Kokon 0.10" update video (Update010, ~41 s, for Reddit, YouTube and the site): Prompt+
(a few words, the prompt written, the image hatching on the drop), "One model. Two jobs." (Prompt+
runs on the Qwen3-VL 8B text encoder Qwen-Image 2.1 already loads), your own LoRAs (beta: Settings,
then the LoRA pill), the new look, the end card. Its clock is src/app/timeline010.ts (beats,
shots, voice cues), its scenes src/scenes/Update010.tsx, its script audio/voiceover-0.10.md.
Until its own takes exist it plays the launch track's tail and no voice (voice-0.10.json holds
estimated spans, marked _provisional).
python audio/gen_music.py update # takes/update-<stamp>-N.wav (ACE-Step API running)
D:/AI/tts/.venv/Scripts/python audio/gen_voice.py audition # voice/update-0.10/audition-<stamp>/
D:/AI/tts/.venv/Scripts/python audio/gen_voice.py build --from audio/voice/update-0.10/audition-<stamp> --pick 1.0:11,2.0:11,...
npx remotion render Update010 out/kokon-0.10-update.mp4 # in video/, 1920×1080
"Kokon 0.11" video (Update011, ~77 s): Kokon on other devices. The site's "Other devices"
scene (src/app/Devices011.tsx: the computer with the GPU, a MacBook and an iPhone, frames from
Devices.css, MIT, in src/app/devices.css), the three setup steps, the phone in use, safe by
default, the server and Docker, the end card. The app on screen is the app itself: Settings →
Other devices in the desktop window, and the browser UI in the phone layout (loaders/phone-container.cjs
turns the 720 px media query into a container query; src/mocks/platform.ts switches desktop and
browser per window). The QR codes in the camera shot and on the end card open the setup guide.
Clock src/app/timeline011.ts (bars of the music, shots, voice cues), script audio/voiceover-0.11.md,
music the launch film's house-4 played whole at one level (faded out over the end card), new
effects make_sfx.py 011 (three variants each in public/sfx/011, the one played is picked in
sound.tsx).
D:/AI/tts/.venv/Scripts/python audio/gen_voice.py audition-011 --seeds 3 # voice/update-0.11/audition-<stamp>/
D:/AI/tts/.venv/Scripts/python audio/gen_voice.py build-011 --from voice/update-0.11/audition-<stamp> --pick 1.0:17,...
npx remotion render Update011 out/kokon-0.11.mp4 # in video/, 1920×1080
"Use cases" video (UseCases, ~87 s, English, narrated, for the site and YouTube): it opens
on why Kokon (private, unlimited), then thirteen everyday edits (photobombers removed, a blink
fixed, an old photo restored and coloured, a missing person added, a haircut and a jacket tried on,
a room tidied, a home redone, a headshot, an invitation with text), each shown as the before photo,
the prompt typed into the prompt box (the app's .composer markup and styles), a wipe to the after
landing on the spoken word, and a caption. Every image was made with Kokon (public/img/usecases/,
web copies of %USERPROFILE%\Pictures\Kokon\use-cases-2k); the video says so on screen. No camera
moves: hard cuts on the beat, wipes and type. Its clock is src/app/timelineUseCases.ts (beat
lengths, moods, where each voice line starts, the word each wipe lands on), its scene
src/scenes/UseCases.tsx, its script audio/voiceover-usecases.md: one narrator with a tone per
section, the launch voice on Chatterbox, and Chatterbox Turbo cloned from it for the playful lines
(with its [chuckle] and [happy] tags). The music is one ACE-Step track with a calm piano part,
cut on bar lines so that part covers the serious beats and the groove returns on the flamingo. The
YouTube thumbnail is UseCasesThumb.
D:/AI/tts/.venv/Scripts/python audio/gen_voice.py audition-usecases --seeds 4 # voice/usecases/audition-<stamp>
D:/AI/tts/.venv/Scripts/python audio/gen_voice.py build-usecases --from audio/voice/usecases/audition-<stamp> --pick 1.0:14,...
python audio/gen_music.py usecases # takes/usecases-<stamp>-N.wav (ACE-Step API running)
D:/AI/tts/.venv/Scripts/python audio/make_update_music.py usecases # video/public/music-usecases.m4a
python audio/make_sfx.py pop swipe deep # only this video's effects
npx remotion render UseCases out/usecases-1.mp4 # in video/, 1920×1080
npx remotion still UseCasesThumb out/usecases-thumb.png # 1280×720
The word each wipe lands on comes from src/app/voice-usecases-words.json (Whisper word times of
the built lines, whisper-small.en): regenerate it after a new build.
src/app/timeline.ts is the film's clock: scene lengths, the frame where each voice-over sentence
plays, the frames of the actions the voice names (Remove, Add, the montage cuts) and the music's
beat grid. Retime there and pictures, voice and music move together.
Sound (src/app/sound.tsx, all levels in one table):
- Voice-over:
public/voice/NN.wav, one file per line ofaudio/voiceover.md, generated locally with Chatterbox (MIT) byaudio/gen_voice.py final, which also writessrc/app/voice.json(where each sentence starts and ends) so sentences can be placed one by one. To change single sentences,gen_voice.py redoandsplicekeep the rest of the approved take (usage at the top of the script). The music ducks about 8 dB under it. - Music:
public/music.m4a, generated locally with ACE-Step 1.5 (MIT, output usable commercially) byaudio/gen_music.py: takehouse-4, 118 BPM, with two bars removed from its breakdown so the second drop (60.3 s) lands on the Agent scene. - Effects:
public/sfx/*.wav, synthesized from scratch byaudio/make_sfx.py(no samples, no licences), kept sparse: typing only in the first demo, quiet clicks, a soft reveal in three variants that take turns when a result appears, whooshes only on big moves.
To regenerate music, start the ACE-Step API (uv run acestep-api in its folder) and run
python audio/gen_music.py house; takes land in audio/takes/ (git-ignored).
The website uses a lighter copy: site/assets/kokon.mp4 (CRF 25, ~17 MB; it autoplays muted and its "Watch with sound" button restarts it with sound) and
site/assets/kokon-poster.jpg. Browsers cache them for a week, so after replacing either one raise the
?v= number where the page links them (site/index.html, site/sitemap.xml, site/assets/style.css;
the same for style.css in every page that links it). To refresh them after a new render:
ffmpeg -y -i marketing/video/out/kokon-launch-0.10.mp4 -c:v libx264 -preset slow -crf 25 -pix_fmt yuv420p -c:a aac -b:a 128k -movflags +faststart site/assets/kokon.mp4
ffmpeg -y -ss 38.5 -i marketing/video/out/kokon-launch-0.10.mp4 -frames:v 1 -q:v 3 site/assets/kokon-poster.jpg
The website
site/index.html plus site/assets/: one static page that links every download button to the
latest GitHub release. Open the file in a browser to preview it. To publish it, any static host
works; the simplest are:
- GitHub Pages (free once the repository is public): Settings → Pages → Deploy from a branch
is limited to
/or/docs, so either movesite/to its owngh-pagesbranch or use a small Pages workflow that uploadssite/. - Cloudflare Pages, Netlify or Vercel: point them at the repository with
siteas the output folder and no build command.