Research map / Image generation / Autoregressive and latent generation
Boosting Resolution Generalization of Diffusion Transformers with Randomized Positional Encodings
Liang Hou, Cong Liu, Mingwu Zheng, Xin Tao, Pengfei Wan, Di Zhang, Kun Gai
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 |
|---|---|---|---|
| add random resize-and-crop augmentation with micro-conditioning (ablation) | RPE-2D DiT-XL/2 · ImageNet 512x512 extrapolation | FID: 20.78 -> 19.12 | Improved |
| add random resize-and-crop augmentation with micro-conditioning (ablation) | RPE-2D DiT-XL/2 · ImageNet 512x512 extrapolation | IS: 293.96 -> 325.76 | Improved |
| add attention scaling at inference (ablation) | RPE-2D DiT-XL/2 · ImageNet 512x512 extrapolation | sFID: 23.91 -> 19.17 | Improved |
| add timestep shifting at inference (ablation) | RPE-2D DiT-XL/2 · ImageNet 512x512 extrapolation | FID: 18.33 -> 17.95 | Improved |
| replace YaRN with RPE-2D for extrapolation 256->512 (ablation) | DiT-XL/2 (RoPE) · ImageNet 512x512 (trained at 256) | sFID: 34.31 -> 18.23 | Improved |
| replace NTK with RPE-2D for extrapolation 512->1024 (ablation) | DiT-XL/2 (RoPE) · ImageNet 1024x1024 (trained at 512) | FID: 32.90 -> 25.40 | Improved |
| apply RPE-2D to RoPE (ablation) | DiT-XL/2 · ImageNet 256x256 in-distribution | FID: 17.30 -> 16.92 | Improved |
| apply RPE-2D to sinusoidal PE (ablation) | DiT-XL/2 · ImageNet 256x256 in-distribution | FID: 17.82 -> 17.33 | Improved |
| compare RPE-2D to PI on precision at 768x768 | DiT-XL/2 · ImageNet 512x512 (trained at 512) | Precision: PI precision 0.8441 vs RPE-2D 0.8271 at 512 | Mixed |
| use RPE-2D when fine-tuning 256->512 | DiT-XL/2 RoPE · 512 fine-tuning | training loss | Improved |
Closest papers by meaning
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.87)
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.87)
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.86)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.86)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.86)
- Recursive Generalization Transformer for Image Super-Resolution (ICLR 2024 · similarity 0.86)
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.
