Research map / Image generation / Text-to-image diffusion
HRDiT: Training-Free High-Resolution Image Generation with Off-the-Shelf Diffusion Transformer Models
Yu Xue, Haoxuan Qu, ZHUOLING Li, Hongbin Xu, Jianxiong Yin, Simon See, Hossein Rahmani, Jun Liu
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 SPA (bundle and slide) (ablation) | HRDiT on FLUX · 4K generation | FID: 75.85 -> 64.91 | Improved |
| add HAP attention pruning (ablation) | HRDiT on FLUX · 4K generation | latency (s): 212 -> 116 | Improved |
| add HAP attention pruning (ablation) | HRDiT on FLUX · 4K generation | FID: 64.82 -> 64.91 | No clear effect |
| add slide operation to bundle (ablation) | SPA · 4K generation FLUX | FID: 69.43 -> 64.91 | Improved |
| use SPA instead of downscaling token indices (ablation) | HRDiT on FLUX · 4K generation | FID: 74.87 -> 64.91 | Improved |
| use head-adaptive scopes instead of fixed local window (ablation) | HAP · 4K generation FLUX | FID: 69.47 -> 64.91 | Improved |
| use head-adaptive scopes instead of random windows (ablation) | HAP · 4K generation FLUX | FID: 73.45 -> 64.91 | Improved |
| estimate scope quality with final-image loss instead of intermediate-feature loss (ablation) | HAP · 4K generation FLUX | FID: 68.02 -> 64.91 | Improved |
| raise attention cost ratio r_c from 0.01 to 0.1 (ablation) | HAP · 4K generation FLUX | FID: 94.51 -> 64.91 | Improved |
| raise attention cost ratio r_c from 0.1 to 0.5 (ablation) | HAP · 4K generation FLUX | latency (s): 116 -> 155 | Hurt |
| increase candidate scopes N_scope from 10 to 50 (ablation) | HAP · 4K generation FLUX | FID: 66.03 -> 64.91 | Improved |
| increase bundle size N from 3 to 5 (ablation) | SPA · 4K generation FLUX | FID: 65.87 -> 64.91 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- HiFlow — HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance (NeurIPS 2025) · better on 9 of 9 shared results
- Direct — DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising (NeurIPS 2026) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.91)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.90)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.90)
- Relay Diffusion: Unifying diffusion process across resolutions for image synthesis (ICLR 2024 · similarity 0.90)
- UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (NeurIPS 2026 · similarity 0.90)
- Native-Resolution Image Synthesis (NeurIPS 2025 · similarity 0.90)
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.
