IC-446In Stable Diffusion v1.5, concept-generating neurons are localized in the second layer of FFNs, spanning less than 3% of FFN parameters, and are disentangled from object-generating neurons

Ruchika Chavhan, Da Li, Timothy Hospedales

SourceConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning

Using WANDA importance scores computed on target-concept versus reference prompts, the authors identified 'skilled neurons' in the second layer of the FFN blocks of SD v1.5's UNet. An ablation across FFN-1, FFN-2, cross-attention value, and CLIP text-encoder layers shows FFN-2 is the most effective pruning target. The skilled neurons span less than 3% of the FFN weight matrix (approximately 0.12% of total model parameters). Pruning these neurons erases the target concept (e.g., Van Gogh style) while preserving the object, whereas pruning unskilled neurons removes the object but retains the style, confirming disentanglement. The authors also note that t-hat = 10 early denoising timesteps suffice, suggesting style and object features form early in the denoising trajectory.

Evidence
interventional
Key metric
less than 3% of FFN weights, approximately 0.12% of total model parameters; FFN-2 outperforms FFN-1, CA-value, and CLIP in erasure (e.g., Van Gogh CLIP similarity 29.2 vs 32.7/32.7/33.2)
Caveat
Erasing specific objects such as 'english springer' inadvertently removes a few related dog breeds, indicating some interference with fine-grained classes.
Model
Stable Diffusion v1.5
Concepts
Depth-dependent structure
Datasets
ImageNetette [eval]
Methods
Wanda [primary]
Related findings
IC-445
Extraction
automatic-extraction