Reproducible · version-pinned

ComfyUI workflows
that actually run on download

Most ComfyUI graphs break the moment someone else downloads them — drifted custom nodes, dead model links, workflows trapped inside stripped PNGs. This repo fixes that: every workflow is pinned to exact node commits and model hashes, with a sample output to verify against. Starting with the LTX chunked-loop video pipeline behind the Nova channel.

MIT-licensed. No private models, no gated LoRAs, no "DM me for the .safetensors."

Why most ComfyUI workflows break on download

A graph that ran perfectly for its author is the most common dead end in local AI. Four reasons, every time.

🧹

Custom-node drift

A node that existed when the graph was saved gets renamed, removed, or replaced by a different author's fork. The workflow won't even load.

📁

Absolute model paths

The creator's models/checkpoints/… path doesn't exist on your machine, so every loader node errors red.

Vague model lists

"You need this LoRA" — with no link, no hash, and a source that may not be public anymore.

📷

Hidden inside a PNG

The workflow lives in image metadata that Civitai, Reddit, and most CDNs strip on upload. You get a picture, not a graph.

What every workflow ships with

Pinned to specific ComfyUI commits and custom-node versions. Real outputs, no broken nodes, no paywalled links you can't access.

📄

The .json, in plain text

Committed as a diffable file you can read and version — not buried inside an image export.

🧾

A README with receipts

Exact custom-node URLs and commit SHAs, every model file with its Hugging Face link and SHA256, and the ComfyUI commit it was tested on.

🎞️

A sample output

The exact image or video that setup produces, so you can confirm visual parity after install instead of guessing.

📥

A drag-and-drop .png

An optional metadata-embedded export for one-drop loading when you'd rather not import JSON by hand.

The workflows

Shipping incrementally as the GPU frees up between Brethof Voice Pro training runs. Each workflow's spec lands first, then its JSON and sample outputs.

Have a workflow worth sharing?

Bring the JSON, the exact custom-node commits, and the model SHA256s — plus a sample output. We test it before it merges. No private models, no gated LoRAs: reproducibility is the whole point.

Contribute on GitHub

Everything we build

External:   YouTube · GitHub