Research map / Image generation / Autoregressive and latent generation
Representation Alignment for Just Image Transformers is not Easier than You Think
Jaeyo Shin, Jiwook Kim, Hyunjung Shim
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 |
|---|---|---|---|
| apply standard REPA (MLP projection) to JiT | JiT-B/16 · ImageNet 256x256 (200 epochs) | FID: 4.37 -> 5.14 | Hurt |
| replace REPA MLP head with Masked Transformer Adapter (PixelREPA) | JiT-B/16 + REPA · ImageNet 256x256 (200 epochs) | FID: 5.14 -> 4.00 | Improved |
| add shallow transformer adapter without masking in place of the MLP head (ablation) | JiT-B/16 + REPA · ImageNet 256x256 (200 epochs) | FID: 5.14 -> 4.68 | Improved |
| add partial token masking (ratio 0.2) to the adapter input (ablation) | PixelREPA adapter · ImageNet 256x256 (200 epochs) | FID: 4.68 -> 4.00 | Improved |
| PixelREPA without masking vs vanilla JiT (ablation) | JiT-B/16 · ImageNet 256x256 (200 epochs) | FID: 4.37 -> 4.68 | Hurt |
| raise mask ratio from 0.2 to 0.5 (ablation) | PixelREPA-B/16 · ImageNet 256x256 (200 epochs) | FID: 4.00 -> 4.58 | Hurt |
| lower mask ratio from 0.2 to 0.1 (ablation) | PixelREPA-B/16 · ImageNet 256x256 (200 epochs) | FID: 4.00 -> 4.26 | Hurt |
| apply PixelREPA to JiT-B/16 | JiT-B/16 · ImageNet 256x256 | FID: 3.66 -> 3.17 | Improved |
| apply PixelREPA to JiT-B/16 | JiT-B/16 · ImageNet 256x256 | Inception Score: 275.1 -> 284.6 | Improved |
| apply PixelREPA to JiT-L/16 | JiT-L/16 · ImageNet 256x256 | FID: 2.36 -> 2.11 | Improved |
| apply PixelREPA to JiT-H/16 | JiT-H/16 · ImageNet 256x256 | FID: 1.86 -> 1.81 | Improved |
| apply REPA to JiT at low resolution | JiT · ImageNet 32x32 | FID | Improved |
Closest papers by meaning
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.92)
- U-REPA: Aligning Diffusion U-Nets to ViTs (NeurIPS 2025 · similarity 0.90)
- AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.90)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.89)
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 2025 · similarity 0.89)
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.
