SharedSAE：跨 U-Net 块与去噪时间的统一稀疏字典
原标题：SharedSAE: a unified sparse dictionary across u-net blocks and denoising time for diffusion models
AI 摘要
研究提出 SharedSAE，一种在 SDXL 1.0 的四个 U-Net 注意力块和全部 50 个去噪步骤间共享的稀疏字典，替代以往按块或按时间步分别训练的多个稀疏自编码器。它通过块特定低秩输入适配器和时间条件模块处理跨块与跨时间的分布差异，重建 FVU 达 0.400，优于逐块方案的 0.492。基于该共享字典的局部特征抑制将 I2P 检测阳性子集中的 NudeNet 阳性生成从 139 降至 1，并揭示概念表征在去噪过程中从早期下采样块向中后期块迁移。
正文节选
SharedSAE: A Unified Sparse Dictionary across U-Net Blocks and Denoising Time for Diffusion Models Abstract Concept representations in diffusion models drift across network depth and denoising time, yet existing sparse autoencoders offer only per-block or per-timestep dictionaries, leaving no single feature dictionary for localization and intervention in the network. We present Shared Sparse Autoencoder (SharedSAE), which learns one sparse dictionary shared across four U-Net attention blocks and