AI 中文总结
针对扩散模型提出重要性感知剪枝框架，通过纳入空间重要性映射到剪枝目标，在MS-COCO数据集上高压缩率下保持主题保真度和结构正确性，证明内容感知目标对生成模型感知忠实压缩很关键。
AI 中文摘要
我们提出了用于扩散模型的重要性感知剪枝，这是一个无需训练的框架，它优先保留对语义显著图像区域至关重要的参数。为此，我们将从条件信号或模型注意力中导出的空间重要性映射纳入剪枝目标。这产生了与感知相关性一致的参数排名，而不是均匀的重建误差。在MS-COCO数据集上，我们提出的方法在高压缩率下始终保持主题保真度和结构正确性，而传统剪枝会导致明显退化。这些结果表明，内容感知目标是生成模型感知忠实压缩的关键。
英文摘要
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.
CommentsAccepted to NeurIPS 2026. Project page: https://sasukepn1999.github.io/importance-obs-pruning/
