FuseReg：正则化层融合缓解表征自编码器的重建-生成差距

AI 摘要

该论文提出 FuseReg，用随机子集采样替代表征自编码器（RAE）中固定的编码器层融合启发式，以缓解重建与生成对层深偏好不同造成的失配。理论分析表明子集采样显式惩罚跨层不一致的敏感性。在 ImageNet-256 与 DINOv3-L 上，单个 FuseReg 解码器无需重训即可从全层、稀疏及单层融合中重建，PSNR 高于固定融合专用解码器；仅替换解码器即让未引导 gFID 降低 27%，联合正则化在 DiT-Base 上降低 29%。

正文节选

FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders Abstract Representation autoencoders (RAEs) reuse features from a pretrained visual encoder as reconstruction and diffusion latents, integrating strong visual representations into image generation. However, RAEs still need to decide which encoder layers form the shared latent space for the generator and pixel decoder. This choice involves a trade-off. Shallower layers tend to preserve fi

发布时间：—

抓取时间：2026-09-28 10:52

来源机构：Hugging Face