Drop an image and get the ComfyUI workflow back: a Civitai link, a PNG with metadata, or your own output. Unbake reads the PNG info — prompt, seed, sampler, steps, CFG, LoRAs — rebuilds the graph, and sorts the missing models by how you would obtain them. Then it runs a validated XY plot (parameter sweep) that throws if anything other than the axis you declared has moved, so the comparison you are looking at is actually a comparison.
INSTALL
ComfyUI-Manager → Custom Nodes Manager → search "Unbake" → Install.
If it does not come up, install by URL: https://github.com/syugoji/ComfyUI-Unbake
Source, README and issues: https://github.com/syugoji/ComfyUI-Unbake
License: GPL-3.0. No dependencies outside ComfyUI itself.
The file attached to this version is not the extension — it is a text pointer to the links above. Civitai requires every version to carry a file. A source copy pinned here would go stale the next time the repository moves, so there is not one.
BEFORE YOU CAN DOWNLOAD MISSING MODELS: SET A CIVITAI API KEY
Reading a record needs no key. Hand it a Civitai image URL and the generation data comes back anonymously — prompt, seed, sampler, steps, CFG, model.
Fetching the model files is a different matter: Civitai requires an API key for that. Open Unbake's settings, paste your key into "Civitai API key", and press Save. Until you do, the download step has nothing to authenticate with.
WHAT THE PICTURES SHOW
The two side-by-side pairs are the same record replayed twice.
The first run was still short one node pack. Unbake asked ComfyUI-Manager to install it, and Manager only registers new nodes at startup — so until ComfyUI was restarted, the graph ran with a substitute. A picture comes out, and it is not the same picture.
The second run is the same record after that restart, with nothing else touched.
Unbake called the first one "medium" and the second one "high" before either was executed. That verdict, and the fact that it moved for a reason you can name, is the point.
WHAT IT DELIBERATELY DOES NOT DO
It does not browse, search or organise models. That is not this tool's job.
It is not a Civitai client. It does nothing beyond being handed one image.
No training, merging or quantising.
Automatic scores in a sweep are shown for reference only. A human picks the winner.
UI is available in 12 languages, including right-to-left (Arabic, Persian).
WATCH IT RUN
Both clips are silent and have no subtitles.
The full walkthrough — install the extension, create the API key, enter it, download what is missing, run it before and after the restart: https://youtu.be/REJc3RicGZs
And the short one, which is only the verdict changing: