Importance-Aware OBS Pruning for Diffusion Models

2026-07-22

AI Summary

This work addresses the degradation in semantic fidelity and structural integrity of images generated by diffusion models under high compression ratios using conventional pruning methods. To mitigate this issue, the authors propose a training-free, importance-aware pruning framework that integrates spatial importance maps—derived from either conditioning signals or attention mechanisms—into the Optimal Brain Surgeon (OBS) pruning objective for the first time. This integration prioritizes the retention of weights critical to semantically salient regions during parameter removal. Evaluated on the MS-COCO dataset, the proposed method significantly outperforms existing pruning strategies, preserving both subject fidelity and structural correctness of generated images even at high compression rates, thereby achieving efficient model compression more aligned with human perception.

Abstract

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.

Problem

• diffusion models
• model pruning
• semantic saliency
• perceptual fidelity
• compression

Innovation

• importance-aware pruning
• diffusion models
• spatial importance maps
• perceptual relevance
• generative model compression