README.md

---
license: cc0-1.0
pretty_name: Model Bending Knowledge Base
tags:
- model-bending
- diffusion
- stable-diffusion
- explainability
- xai-for-the-arts
- comfyui
configs:
- config_name: default
  data_files: index/viewer*.parquet
---

# Model Bending Knowledge Base

This dataset records what happens when you bend the inside of a diffusion model. Bending means multiplying, rotating,
adding noise to or otherwise changing the activations of a layer while the model generates.

Each record names:
- the model and the exact part of it that was bent
- the operation, the amount, and the denoising steps it covered
- the full generation setup
- the output, next to an unbent baseline made with the same setup

Artists can browse it to learn what a model does when bent. Agents, such as the
[comfyui-model-bending](https://github.com/abuzreq/comfyui-model-bending-agent) skill, query it to suggest starting
recipes ("more abstract on SD1.5" → which bends tend to do that).

Bends are applied with [ComfyUI-Model-Bending](https://github.com/abuzreq/ComfyUI-Model-Bending).

**17454 records**, 1715 cells and 10 findings (index built 2026-10-06T00:47:55Z). Sources: author_experiment 2940, paper 5656, sweep 8858. Model families: sd1 17454.

## Facts are kept apart from interpretation

Every record folder `records/<family>/<source>/<id>/` holds:

| file | what | who made it |
|---|---|---|
| `record.json` | **facts**: model, checkpoint, sampler, scheduler, steps, cfg, seed, size, route; the bends as actually applied (layer path, op, arguments, step window); the output file | the producer named in `provenance` |
| `measurements.json` | **numbers** computed against the unbent baseline: MAE, latent cosine distance, LPIPS, DINOv2 and CLIP distances, prompt retention, a CLIP degeneracy score, pixel flags, and whether the render broke | each value names its method, and its model when a learned model computed it |
| `interpretations.jsonl` | **interpretation**: captions, "what changed", keywords, effect tags, concept tags, notes, verdicts | each line names its author: a human (`{"type": "human", "name": …}`) or an AI (`{"type": "ai", "model": <exact model id>, "prompt_version": …}`) |
| `output.webp` | the output image | |

Other folders:
- **`baselines/`** holds the unbent renders.
- **`findings/`** holds claims about many records, such as a paper's results, with their authors and citation:
  - `level: cell` findings back the cells they cover.
  - `level: general` findings give study-wide context.
- **`vocab/effects.json`** is the controlled list of effect tags.
- **`schema/`** holds the JSON Schemas.

Prompts and input images are published only when their owner agreed (`consent`). Otherwise a salted key stands in, so
records can still be counted per prompt.

AI-written interpretations are always labelled with the model that wrote them. Treat them as one reading of the image,
not ground truth. Humans can add their own readings next to them.

### Run a record in ComfyUI

Every record has a `workflow.json`, and the same workflow is embedded in its `output.webp`:
- **Drag the image onto ComfyUI** to open a graph that renders it: the checkpoint and any separate loaders, the
  sampler, steps, cfg, seed, size, prompt and negative, and the bend in an **Apply Bends from JSON** node
  ([ComfyUI-Model-Bending](https://github.com/abuzreq/ComfyUI-Model-Bending)) with clamping off.
- **Prompts that were not shared** appear as a placeholder; put in your own.
- **Expect a close match, not a pixel-exact one:** renders can differ slightly across ComfyUI and torch versions and
  GPUs.

### Stable ids

A record id is a hash of a fixed list of facts (`id_scheme`):
- model: family, arch, checkpoint
- setup: route, seed, sampler, scheduler, steps, cfg, size, denoise, prompt (or its key), negative, input key
- bends: path, op, arguments, step window, blend

Adding records or new fields never changes a published id. Ids from before a migration resolve through
`migrations/id_scheme_*.json`, published as `index/id_aliases.json`.

An absent `negative` means it was not recorded; `""` means it was empty.

## Cells and evidence

`index/cells.jsonl` groups single-bend records into **cells**: (model family, layer group, sub-module kind, module type,
op, amount bucket, step window, route). Each cell carries:
- record, seed and prompt counts
- the checkpoints it was tested on
- measurement summaries
- effect tags, with who assigned them
- an evidence grade:
  - anecdotal: one seed and one prompt
  - multi-seed or multi-prompt
  - replicated: at least two of each
  - study-backed: a cited finding covers the cell
- what can go wrong (see below)

### Broken renders

A render counts as **broken** (`degenerate`) only when it clearly failed:
- CLIP ViT-B/32 reads it as noise or a blob field (`clip_degenerate` ≥ 0.90) and it lost part of its subject
  (`prompt_retention` < 0.85, CLIP's image–prompt match relative to the unbent baseline)
- or it is static-like noise, or an all-black or all-white frame

These cut-offs were set from the dataset owner's labels on 59 renders near the boundary, so that no image they judged
usable is marked broken. The cost is that some failures go unflagged. Losing the subject, flat textures and smooth
blob fields are recorded (`prompt_retention`, `pixel_flags`) but do not count as broken on their own: artists often
want them.

In `index/cells.jsonl`, broken renders are left out of a cell's measurement summaries, effect tags and examples. Each
cell still reports what can go wrong:
- `degenerate_rate` and `broken_reasons`: how often its bend broke, and why
- `signals`: warning signs over all renders:
  - `subject_fades`: the share with prompt retention < 0.75
  - `noise_or_blob_look`: the share with a CLIP degeneracy score ≥ 0.6
- `failure_notes`: what failing renders looked like, with their author
- `broken_examples`: the broken records themselves

## Sources

| source | what |
|---|---|
| `paper` | Experiments from *Unboxing Diffusion Models for the Arts: Interactive Model Bending and Practice-Based Explainability*: RealisticVision v5.1 (SD1.5), all-layer multiply and noise sweeps, and the multi-seed, multi-prompt and timestep studies |
| `author_experiment` | Further sweeps by the same author, e.g. rotation across 14 prompts on SD1.4 |
| `sweep` | Systematic runs: the maintainer's graded sweeps, and full runs contributed by others. Contributed ones carry a `contribution` block (who, by script or agent, the run, the CC0 agreement). |

## Contributing a run

The knowledge base takes **full runs**, not single pictures:
- one model
- all seven regions of its U-Net
- at least three operations with three amounts each

One prompt and seed is enough.

Make a run with the bending skill's
[`kb_run.py`](https://github.com/abuzreq/comfyui-model-bending-agent/blob/main/skills/comfyui-model-bending/scripts/kb_run.py)
(`init` → `render` → `check` → `submit`). It renders on your own ComfyUI, embeds each picture's workflow, checks the
format and the coverage, and opens a **Pull Request** here with your own Hugging Face login. Measurements and
descriptions are optional; the maintainer adds whatever a run leaves out. Every interpretation must name its author,
and AI-written ones must name their model. The steps are in the
[Navigator](https://abuzreq-model-bending-navigator.static.hf.space/?page=contribute).

## Requesting a model

Ask for a model in this dataset's **Discussions**: a discussion titled `Model request: <name>`. The pinned post
explains what to include (link, licence, settings, 3–5 test prompts with negatives). Vote with 👍. The most wanted
models get a full run.

## Likes

People signed in to Hugging Face can like renders in the Navigator, once per account and render. Only the counts are
published, in `community/likes.jsonl` (`{record, likes}`), by the
likes server (a Netlify function); who liked what is kept privately.
Cells in `index/cells.jsonl` sum their renders' likes as `likes`.

## Citation

```bibtex
@misc{abuzuraiq2026unboxing,
  title  = {Unboxing Diffusion Models for the Arts: Interactive Model Bending and Practice-Based Explainability},
  author = {Abuzuraiq, Ahmed M. and Pasquier, Philippe},
  year   = {2026},
  eprint = {2607.22428},
  archivePrefix = {arXiv}
}
```