https://github.com/user-attachments/assets/751d01c0-2dc6-427d-b00d-ec5d7f2c0249

nd-video-studio

Explainer videos in HTML: compositions in HyperFrames
(Apache 2.0) → MP4, plus narrator, music and mixing on NeuralDeep infrastructure.
The whole pipeline is described in the skills/nd-video/SKILL.md skill,
picked up by Claude Code (.claude/skills) and Codex/Cursor (.agents/skills).

Example: LLM vs Jev

The video at the top (media/jev-vs-llm.mp4, 1080p, 112 s, 8.7 MB) is the first one made with this
pipeline: a camera flying over two architecture "crystals", a magnifier for details, word-level
subtitles, narrator with the trailer preset and ducked music.

What's inside

path	what
`skills/nd-video/`	skill: from idea to MP4 (HyperFrames techniques, voice, music, mixing)
`skills/ndt-content/`	skill: a video from a brief entirely on hub models (research, script, voice, images, checks, review)
`scripts/nd_probe.py`	availability check of every hub model and endpoint the pipeline uses
`scripts/nd_research.py`	research via the hub Search API → `sources.md`
`scripts/nd_script.py`	script by a hub model: brief + sources → `narration.json` with scenes
`scripts/nd_retime.py`	line starts from the real voiceover, before laying out the HTML
`scripts/nd_align.py`	`whisper-1`: voiceover vs text check and word timings `words.json`
`scripts/nd_images.py`	hub FLUX images, background removal, upscale
`scripts/nd_review.py`	contact-sheet review by a hub vision model
`scripts/music_studio.py` + `tools/music-studio.html`	local music studio on top of self-hosted ACE-Step: presets, variants, history, "to project"
`scripts/beat_map.py` + `tools/beat-sync.js`	track beat map (beats, downbeats, hits) and binding animation to it, time warp
`scripts/tts_gpt_audio.py`	narrator via OpenAI gpt-audio (OpenRouter), with a verbatim check
`scripts/tts_hub.py`	narrator via the hub TTS (`/v1/audio/speech`)
`scripts/music_lyria.py`	music via Google Lyria 3 (OpenRouter)
`scripts/music_acestep.py`	music via self-hosted ACE-Step 1.5 (docs/acestep.md)
`scripts/fit_narration.py`	sync: lines into timeline windows, silence trimming, atempo
`scripts/voice_fx.py`	EQ, compression, convolution reverb; presets and A/B
`scripts/mix.py`	voice and music mixed with ducking into one wav
`scripts/build_audio.sh`	fit → fx → mix in one command
`scripts/sync_hyperframes.sh`	update the vendored HyperFrames skills
`vendor/hyperframes/`	vendored HyperFrames skills and docs + their LICENSE (NOTICE)

Quick start

mkdir -p projects/my-video && cd projects/my-video    # projects/ are local, not tracked in git
env -u HTTPS_PROXY -u HTTP_PROXY -u https_proxy -u http_proxy -u ALL_PROXY npx -y [email protected] init . --example blank
# narration.json → voice (tts_hub.py) → build_audio.sh → render; step by step in skills/ndt-content/SKILL.md
bash ../../scripts/build_audio.sh . trailer
env -u HTTPS_PROXY -u HTTP_PROXY -u https_proxy -u http_proxy -u ALL_PROXY npx -y [email protected] render --output renders/my-video.mp4

Requires Node 22+, FFmpeg, Python 3.10+ (standard library only).

Keys

OPENROUTER_API_KEY (gpt-audio, Lyria; OpenRouter returns 403 from Russian IPs), ND_API_KEY (hub),
ACESTEP_API_KEY (if the ACE-Step server is key-protected). Keep them in .env, it's in .gitignore.

License

MIT. vendor/hyperframes/ stays under its own Apache License 2.0 (LICENSE).