Research map / Image generation / Autoregressive and latent generation
Residual Connections Harm Generative Representation Learning
Xiao Zhang, Ruoxi Jiang, William Gao, Rebecca Willet, Michael Maire
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decay identity shortcut weights over encoder depth in MAE | MAE ViT-B/16 · ImageNet-1K | linear probing accuracy (%): 67.8 -> 72.7 | Improved |
| decay identity shortcut weights over encoder depth in MAE | MAE ViT-B/16 · ImageNet-1K | KNN accuracy (%): 27.4 -> 63.9 | Improved |
| decay identity shortcut weights in MAE | MAE ViT-B/16 · ImageNet-1K | fine-tuning accuracy (%): 83.6 -> 82.9 | Hurt |
| decay identity shortcuts in MAE | MAE ViT-B/16 · COCO val unsupervised segmentation | mIoU: 4.1 -> 10.4 | Improved |
| add encoder-decoder skip connections alongside decay (ablation) | decayed MAE · ImageNet-100 | linear probing accuracy (%): 61.5 -> 83.6 | Improved |
| use fixed linear schedule instead of cosine schedule (ablation) | decayed MAE wmin=0.6 · ImageNet-100 | linear probing accuracy (%): 82.8 -> 83.6 | Improved |
| use predefined fixed decay instead of learnable shortcut weights (ablation) | decayed MAE · ImageNet-100 | linear probing accuracy (%): 79.5 -> 83.6 | Improved |
| scale only the block output branch f instead of down-weighting shortcut (ablation) | MAE baseline · ImageNet-100 | linear probing accuracy (%): 76.5 -> 76.9 | No clear effect |
| scale both branches by constant sqrt(0.5) (ablation) | MAE baseline · ImageNet-100 | linear probing accuracy (%): 76.5 -> 82.6 | Improved |
| apply decay to both attention and MLP rather than only one (ablation) | decayed MAE · ImageNet-100 | linear probing accuracy (%): 79.3 (attention only) / 80.6 (MLP only) -> 83.6 | Improved |
| decay shortcuts in SiT-XL/2 diffusion with wmin=0.8 | SiT-XL/2 · ImageNet-1K 256x256 class-conditional | FID: 17.2 -> 11.8 at 400k iterations | Improved |
| add decay on top of encoder-decoder skip connections in SiT-XL/2 (ablation) | SiT-XL/2 · ImageNet-1K 256x256 class-conditional | FID: 16.5 -> 11.8 | Improved |
Closest papers by meaning
- Neural Residual Diffusion Models for Deep Scalable Vision Generation (NeurIPS 2024 · similarity 0.87)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.86)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.86)
- Revisiting Residual Connections: Orthogonal Updates for Stable and Efficient Deep Networks (NeurIPS 2025 · similarity 0.85)
- Hyper-Connections (ICLR 2025 · similarity 0.85)
- Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
