Model bending: open-source ComfyUI toolkit for glitching diffusion models mid-sampling

abuzreq · reddit · 2026-10-16

Model bending means reaching into a diffusion model mid-sampling and manipulating its activations: rotating a feature map, zeroing a block, pouring noise into an attention layer. It's neither prompting nor a style LoRA — closer to circuit bending: break the model on purpose and see what falls out. With the same prompt, seed and model, bending a single layer turns a fisherman into a dockside scene, a burlap twin, etchings or leopard print. Each effect needs a recipe: model + layer address + bending operation/amount.

Everything is open source:

Model Bending Navigator (HF Space): 17,000+ bending examples, each a single bend next to its unbent original; filter by effect, model part or operation, then hit Run this bend to get a ready-to-run ComfyUI workflow.
ComfyUI Model Bending (MIT, in ComfyUI Manager): bend any layer with add/multiply/noise/rotate/scale/blur/Sobel/Fourier amplify and more; interactive web UI to inspect the model, click a layer, drag a slider; limit bends to early/middle/late steps (early rebuilds composition, late messes with surface); supports SD1.5/SDXL U-Nets plus Flux, SD3 and other DiTs, and can bend LoRAs and VAEs too.
Experimental WAN 2.1/2.2 video bending (only tested on tiny random-weight models so far).

The project comes out of the author's PhD research on treating AI models as a material for art, aiming to reintroduce surprise into generative AI.

Original post →

Model bending: open-source ComfyUI toolkit for glitching diffusion models mid-sampling

More from Multimodal