Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces an importance-aware pruning method for diffusion models that focuses on preserving parameters crucial to semantically important image regions. By utilizing spatial importance maps derived from conditioning signals or model attention, the approach achieves parameter rankings that prioritize perceptual relevance over standard reconstruction error. The method shows significant improvements in subject fidelity and structural integrity at high compression ratios on the MS-COCO dataset, highlighting the effectiveness of content-aware objectives in maintaining the quality of generative models during compression.
Pruning diffusion models with a focus on perceptual relevance can maintain image fidelity even at extreme compression rates, challenging traditional methods that prioritize uniform reconstruction error.
We propose importance-aware pruning for diffusion models, a training-free framework that prioritizes preserving parameters critical to semantically salient image regions. To do so, we incorporate spatial importance maps -- derived from conditioning signals or model attention -- into the pruning objective. This produces parameter rankings aligned with perceptual relevance rather than uniform reconstruction error. On MS-COCO dataset, our proposed approach consistently retains subject fidelity and structural correctness at high compression ratios where conventional pruning causes visible degradation. These results demonstrate that content-aware objectives are key to perceptually faithful compression of generative models.