Datasets:
Read first: how to request a model
This is where you ask for models to be added to the Model Bending Knowledge Base, and vote for the ones you want
most. The most wanted models get a full run: one bend at a time, across every part of the model, with several
operations and amounts, all measured and described. Browse what is already there in the
Model Bending Navigator.
Before you ask
- Search the open requests first. If your model is already there, vote for it with 👍 instead of opening a new
one. - One model per request. Title it
Model request: <name>, e.g.Model request: DreamShaper 8. - The licence must allow sharing renders under CC0. Everything in the knowledge base is CC0.
- Models the bending nodes can reach: SD 1.x, SD 2 and SDXL, including fine-tunes, and the transformer models Flux.1, Flux.2, SD3 / SD3.5 and WAN (video). For something else, ask anyway: requests help decide what comes next
What to write
Copy this and fill it in:
**Model:** <link on huggingface.co or civitai.com>
**Family:** SD 1.x / SD 2.x / SDXL / SD 3 / Flux / other
**Licence:** <it must allow sharing renders under CC0>
**Settings:** sampler, scheduler, steps, cfg, size (and any VAE or LoRA)
**Test prompts** (3 to 5, each with an optional negative):
1. prompt: … | negative: …
2. prompt: … | negative: …
3. prompt: … | negative: …
**Why this model:** …
Test prompts should show what the model is good at. Don't name living artists, brands, film titles or private
people.
Voting
React with 👍 to a request's first post. Only the number of votes counts, and requests are sorted by it in the
Navigator.
What happens next
The maintainer replies to each request. When a model has been run and is in the knowledge base, the request is
closed with a link to it. If it can't be done (licence, size, missing support), it is closed with the reason.
Can't wait?
Run it yourself. A full run takes about 10–30 minutes on a mid-range GPU with the bending skill's kb_run.py, and
you send it as a pull request. The steps are on the
Contribute page.