Research map / Label-efficient and robust learning / Recognition and long-tail
What matters for Representation Alignment: Global Information or Spatial Structure?
Jaskirat Singh, Xingjian Leng, Zongze Wu, Liang Zheng, Richard Zhang, Eli Shechtman, Saining Xie
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 |
|---|---|---|---|
| use a target encoder with higher ImageNet-1K accuracy (PE-Core-G vs PE-Spatial-B) | REPA · ImageNet 256 generation, SiT-XL/2, 100K | FID: FID 32.3 (82.8% acc) vs 21.0 (53.1% acc) | Hurt |
| use larger encoder within same family (WebSSL-1B higher accuracy) | REPA · ImageNet 256 generation, SiT-XL/2, 100K | FID: FID 26.1 vs 21.0 (acc 76.0% vs 53.1%) | Hurt |
| inject global info into patch tokens via CLS mixing (alpha 0 to 0.5) (ablation) | REPA with DINOv2 target · ImageNet 256 generation, SiT-XL/2 | FID: 19.2 -> 25.4 (linear probe 70.7% -> 78.5%) | Hurt |
| replace MLP projection with 3x3 convolution and add spatial normalization (iREPA) | REPA, SiT-XL/2 DINOv2-B · ImageNet 256, 100K steps, no CFG | FID: 19.06 -> 16.96 | Improved |
| add iREPA to weak-spatial-structure target PE-Core-G | REPA, SiT-XL/2 · ImageNet 256, 100K steps, no CFG | FID: 32.35 -> 18.19 | Improved |
| add spatial normalization layer (to conv projection) (ablation) | iREPA · ImageNet 256, 100K steps, SiT-XL/2, DINOv2-B | FID: 18.52 -> 16.96 | Improved |
| add convolution projection (to spatial normalization) (ablation) | iREPA · ImageNet 256, 100K steps, SiT-XL/2, DINOv2-B | FID: 17.66 -> 16.96 | Improved |
| add iREPA across encoder sizes PE-B, PE-L, PE-G | REPA, SiT-XL/2 · ImageNet 256, 100K steps | FID: PE-B 22.5 -> 17.5; PE-L 28.7 -> 17.6; PE-G 32.3 -> 19.5 | Improved |
| add iREPA across diffusion model sizes SiT-B, L, XL | REPA, DINOv2 target · ImageNet 256, 100K steps | FID: SiT-B 49.50 -> 43.37; SiT-L 24.10 -> 20.28; SiT-XL 19.06 -> 16.96 | Improved |
| add iREPA at alignment layers 4, 6, 8 | REPA, SiT-B/2 · ImageNet 256, 100K steps | FID: Layer 4: 49.0 -> 36.0; Layer 8: 54.8 -> 38.9 | Improved |
| add iREPA to REPA-E | REPA-E, SiT-XL/2 · ImageNet 256, 100K steps | FID: WebSSL-1B 26.5 -> 13.2; DINOv2-B 12.9 -> 12.1 | Improved |
| add iREPA to MeanFlow w/ REPA | MeanFlow w/ REPA, SiT-B/2 · ImageNet 256, 100K steps, 1 NFE w/ CFG | FID: DINOv3-B 22.55 -> 17.32 | Improved |
Closest papers by meaning
- When does perceptual alignment benefit vision representations? (NeurIPS 2024 · similarity 0.88)
- Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot (NeurIPS 2026 · similarity 0.88)
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.87)
- Do Vision and Language Encoders Represent the World Similarly? (CVPR 2024 · similarity 0.87)
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.87)
- Discriminating image representations with principal distortions (ICLR 2025 · similarity 0.87)
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.
