让扩散模型剪枝更懂图像重点，压缩后仍保关键细节
Importance-Aware OBS Pruning for Diffusion Models
- 根据图像重要区域动态排序参数，剪掉不重要的部分
- 在高压缩比下保持主体清晰度和结构正确性，无明显失真
- 适合追求生成质量的模型压缩场景，尤其对视觉敏感任务
我们提出一种针对扩散模型的重要性感知剪枝方法，这是一种无需训练的框架，旨在保留与语义显著图像区域相关的关键参数。为此，我们引入空间重要性图——由条件信号或模型注意力生成——并将其融入剪枝目标中，使参数排序更符合感知相关性，而非均匀重建误差。在 MS-COCO 数据集上，该方法在高压缩比下始终维持主体保真度与结构正确性，而传统剪枝方法则出现明显退化。结果表明，内容感知的目标对生成模型的感知保真压缩至关重要。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。