MADHOUSE

A self-hosted pipeline that writes, renders and publishes hour-long music videos.

One recipe goes in. What comes out is an album of original songs, its own cover art, a shader-rendered video, a description with real timestamps, and a slot on the publishing calendar. Nobody sits in the middle of it.

Songs per album	22	generated per run
Output	1920 × 1080 · H.264	≈ 1 hour, AAC 256k
Render rate	197 fps	preset 04, NVENC
Shader presets	4	hand-written GLSL
Publishing pace	2 / day / channel	scheduled automatically
Hosts	1	private network only

Preset 04 — spectrum, 56 bands06 · After Hours Pressure

20 Hz2001k5k20 kHz

Tracklist

04Locking Up the Warehouse
05Night Shift Groove
06After Hours Pressure
07Warehouse After Hours
08Late Shift Grooves

Eight stages, each fed by the one before it

Every stage is resumable on its own. A service restart, a rate limit, or a provider returning an ambiguous 500 does not cost the album — the daemon resumes where it stopped rather than starting over.

Stage	Produces	Worked by
01 Recipe	Genre, mood, market, language, song count Saved per channel and reused for every run on it.	operator
02 Songs	22 original tracks with lyrics One job per slot, sent in parallel; failed slots retry without losing finished ones.	Lyria / useapi
03 Cover	One 2K album cover Upscaled from one of the album's own song covers, so the video's face comes from its own music.	Google Flow
04 Titles	Album and track titles Written against the target market, deduplicated against everything the channel already published.	LLM gateway
05 Mix	One continuous audio file Only the ticked songs, in order, with timestamps computed from real durations.	ffmpeg
06 Render	The finished video A shader scene reads the spectrogram and draws each frame on the GPU. Preview and export share one render spec.	OpenGL · NVENC
07 Publish	A scheduled, described video The upload is matched to its album by title; schedule and tracklist description travel in one API call.	YouTube Data API
08 Playlist	Position in the channel playlist Added oldest first, so the order on the page matches the order it was made.	YouTube Data API

Four scenes, written by hand in GLSL

No stock templates and no video editor. Each frame is drawn from the audio itself: the spectrogram, the running volume, and the standard deviation that spikes on a snare.

Preset	Treatment	Technique
01 Visualizer	Radial spectrum around a circular hole The cover breathes behind it; blur only appears when the track gets loud.	polar · waveform
02 Music Bars	Linear stereo spectrum Left channel below the axis, right above, over a dimmed cover.	linear · stereo
03 Motion Trails	Reactive ring over its own past frames Movement leaves a tail; the only preset that keeps state between frames.	temporal · multi-pass
04 Vinyl Pulse	Full-screen cover on a heartbeat envelope Sharp mini cover, title, spectrum and the full tracklist sit on top, with the playing track in red.	atlas · per-song art

Measurements, with the method attached

Nothing here is an estimate. Frame rates were sampled on the machine under real load; quota figures are counted per call by the application itself against Google's published table.

Figure	Value	How it was obtained
Encoder gain	2.5 ×	NVENC 235 fps against x264 95 fps, same scene, same source, same machine.
Render rate	197 fps	Preset 04 at 1080p with tracklist and per-song artwork drawn, 180-frame run.
Scheduling cost	51 units	One read plus one write per video. Description rides the same write, which is why it is 51 and not 102.
Playlist cost	50 units	Per video inserted — as expensive as scheduling it, so inserted videos are recorded and never re-sent.

One machine, and it is not on the internet

What it runs on

A single Linux host with an NVIDIA GPU. A FastAPI service owns the data, a separate daemon owns the render queue and the pipeline tick, and a React dashboard drives both.

SQLite per workspace in WAL mode — no database server to operate
The render daemon is its own process, so restarting the web app cannot strand a job
Every stage records where it stopped, so a restart resumes instead of repeating

Reachable only from its own network

The dashboard and the API are published to a private network and nothing else. There is no public port, no shared tenancy, and no third party holding the data.

Google credentials never leave the machine and never reach the browser
Uploads are done by hand in YouTube Studio — the app never uploads for you
The privacy policy lists exactly what is stored

Built for one operator, on purpose

MADHOUSE is Clawfog's own production system, not a product you sign up for. If the approach is useful to you, say so and we can talk.

hello@clawfog.com →