A Stable Diffusion-style U-Net rebuilt with B-cos layers produces per-token pixel attributions, and those attributions can flag prompts whose concepts are missing from the generated image.
B-cos networks: Alignment is all we need for interpretability
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Interpretable Diffusion Models with B-cos Networks
A Stable Diffusion-style U-Net rebuilt with B-cos layers produces per-token pixel attributions, and those attributions can flag prompts whose concepts are missing from the generated image.