AI
FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 24/09/2026, 20:00:00 UTC
Briefing created: 28/09/2026, 02:19:24 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/09/2026, 02:18:25 UTC.
Briefings
FuseReg addresses the reconstruction-generation gap in representation autoencoders by replacing heuristic layer selection with training over random encoder subsets. Theoretical analysis shows this penalizes sensitivity to cross-layer disagreement, improving PSNR on ImageNet-256 with DINOv3-L compared to fixed-fusion decoders. Additionally, decoder replacement alone reduced unguided gFID by 27% on RAEv2 DiT-XL, and joint regularization of both stages lowered gFID by 29% on DiT-Base diffusion models.
Key points
- FuseReg replaces heuristic feature selection with training over random subsets of encoder layers to mitigate the reconstruction-generation gap.
- On ImageNet-256 with DINOv3-L, FuseReg decoders achieved higher PSNR than specialized decoders using fixed layer fusion strategies.
Context and limitations
Results are limited to ImageNet-256 and DiT-XL/Base architectures; no performance data provided for other datasets or model sizes.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 28/09/2026, 02:19:24 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiaftcovoi5y7r4fmmndmwcapbclm5xe3sjfpactqpymwvrgw467v4