wooyeolbaek/attention-map-diffusers

Enables extraction and visualization of attention weights across diffusion model layers during inference, supporting both UNet and DiT-based architectures across Stable Diffusion 2/3/3.5, FLUX, and other Diffusers models. Uses module replacement and forward hooks to capture cross-attention tensors at configurable timesteps and layers, then maps attention weights back to tokenized prompt text for per-token spatial visualization. Integrates directly with HuggingFace Diffusers pipelines through a simple `init_pipeline()` call, requiring minimal code changes to existing inference workflows.

397 stars and 87 monthly downloads. Available on PyPI.