2.9K ·GitHub
What is this?
A generative, code-rendered music video for the song 'I'm Upping My P(doom)' with word-synced karaoke-style typography, where every frame is a deterministic function of song time.
A generative, code-rendered music video for the song 'I'm Upping My P(doom)' with word-synced karaoke-style typography, where every frame is a deterministic function of song time.Is this a finished video file or a renderer?It's a TypeScript + three.js renderer app, not a pre-rendered video file: the same code drives both the live browser preview and the offline 1080p60 or 4K60 export, so they look identical.
It's a TypeScript + three.js renderer app, not a pre-rendered video file: the same code drives both the live browser preview and the offline 1080p60 or 4K60 export, so they look identical.How do I preview or render the video myself?Run the bun + Vite app to get a real-time interactive browser preview, then use scripts/render.ts to export offline — it drives headless Chrome over WebSocket into ffmpeg.
Run the bun + Vite app to get a real-time interactive browser preview, then use scripts/render.ts to export offline — it drives headless Chrome over WebSocket into ffmpeg.What do I need installed first?You need bun, Google Chrome (driven headless via playwright-core) and ffmpeg with libx264; regenerating the lyric/audio timing data additionally needs uv for the Python analysis tools.
You need bun, Google Chrome (driven headless via playwright-core) and ffmpeg with libx264; regenerating the lyric/audio timing data additionally needs uv for the Python analysis tools.Does the export run in real time like the preview?No — the live preview runs in real time on a recent Mac, but the offline export is much heavier: rendering the whole song in 4K took about 2.5 hours on an M5 Pro.
No — the live preview runs in real time on a recent Mac, but the offline export is much heavier: rendering the whole song in 4K took about 2.5 hours on an M5 Pro.What exactly does the MIT license cover here?Only the renderer code is MIT-licensed; the bundled fonts keep their own licenses, and the song and lyrics belong to their original authors and are excluded from the MIT grant.
Only the renderer code is MIT-licensed; the bundled fonts keep their own licenses, and the song and lyrics belong to their original authors and are excluded from the MIT grant.How GPU-intensive is a 4K render?Very — it's GPU-bound, with per-frame cost ranging from about 40ms for a still frame to over 10 seconds for ray-marched rooms at high sub-frame counts, and each render pipeline uses roughly 5GB for headless Chrome plus 4GB for ffmpeg.
Very — it's GPU-bound, with per-frame cost ranging from about 40ms for a still frame to over 10 seconds for ray-marched rooms at high sub-frame counts, and each render pipeline uses roughly 5GB for headless Chrome plus 4GB for ffmpeg.What determines how much motion-blur sampling a frame gets?The renderer's --samples auto mode adaptively picks between 12 and 324 sub-frames per frame depending on how fast things are moving, stopping once more samples wouldn't change the image beyond a set tolerance.
The renderer's --samples auto mode adaptively picks between 12 and 324 sub-frames per frame depending on how fast things are moving, stopping once more samples wouldn't change the image beyond a set tolerance.How was the word-level lyric timing created?Python tools run via uv use Demucs for stem separation, CTC forced alignment cross-checked with Whisper, and beat/onset analysis to produce the committed data/lyrics.json and data/audio.json timing files.
Python tools run via uv use Demucs for stem separation, CTC forced alignment cross-checked with Whisper, and beat/onset analysis to produce the committed data/lyrics.json and data/audio.json timing files.Why does the YouTube upload look different from a local 4K render?The YouTube version only averages 4 sub-frames per frame for motion blur and gets recompressed, so fast motion looks stepped and the film grain is smeared, while a local render can use up to 324 sub-frames and keeps full detail.
The YouTube version only averages 4 sub-frames per frame for motion blur and gets recompressed, so fast motion looks stepped and the film grain is smeared, while a local render can use up to 324 sub-frames and keeps full detail.What role did Claude play in making the video?Claude (Opus 5.5) in Claude Code was used in conversation with the author to work out the concept, lyric alignment, renderer code, every scene, and the final renders.
Claude (Opus 5.5) in Claude Code was used in conversation with the author to work out the concept, lyric alignment, renderer code, every scene, and the final renders.What are similar projects?ClaudeAnimationBase, mg-styles-15, huashu-art-motion, awesome-opus5-5-videos, riso-windowseat, saas-motion-kit, dioramas, lemo-opuscar
ClaudeAnimationBase, mg-styles-15, huashu-art-motion, awesome-opus5-5-videos, riso-windowseat, saas-motion-kit, dioramas, lemo-opuscarWhich topics is it in?procedural, headless-browser, motion-graphics, generative-art, creative-coding, audio-processing, whisper, typography, threejs, webgl, bun, music
procedural, headless-browser, motion-graphics, generative-art, creative-coding, audio-processing, whisper, typography, threejs, webgl, bun, music