Research map / Image generation / Autoregressive and latent generation
UltraGen: Efficient Ultra-High-Resolution Image Generation with Hierarchical Local Attention
Yuyao Zhang, Yu-Wing Tai
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 hierarchical local-window attention with low-res guidance instead of dense FlashAttention at 8K | FLUX.1-dev · 8K generation | total inference speedup: 10.77x total, 6.93x per iteration | Improved |
| use window-first local attention instead of naive FlashAttention | FLUX.1-dev · 4K generation | additional VRAM: 15.6 GB -> 7.3 GB | Improved |
| generate at 2K-4K with model trained only on 1K data | UltraImageGen · GenEval | overall: 0.66 -> 0.67 | No clear effect |
| let high-res tokens attend to low-res guidance tokens (ablation) | UltraImageGen · qualitative 4K | artifacts | Improved |
| allow each window to attend to neighboring windows (ablation) | UltraImageGen · qualitative 4K | boundary artifacts: under 2% extra compute | Improved |
| vary window size (256/512/1024) (ablation) | UltraImageGen · 4K | FID: 67.23 / 66.31 / 67.83 | No clear effect |
| change guidance ratio from 4 to 2 (ablation) | UltraImageGen · 4K | FID: 67.23 -> 67.19 | No clear effect |
| increase LoRA rank from 4 to 16 (ablation) | UltraImageGen · FID at 10K steps | FID: 82.41 -> 76.08 | Improved |
| increase LoRA rank from 16 to 64 (ablation) | UltraImageGen · FID at 10K steps | FID: 76.08 -> 76.15 | No clear effect |
| skip more denoising steps from low-res warm start (ablation) | UltraImageGen · 4K | PSNR vs full-step: 28.78 (4 steps) -> 26.43 (18 steps), speedup 1.17x -> 2.8x | Mixed |
| switch backbone to FLUX.2 | UltraImageGen · 4K | FID: 67.03 -> 61.38 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- HiFlow — HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance (NeurIPS 2025) · better on 1 of 1 shared result
- UltraPixel — UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024) · better on 0 of 1 shared result
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
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.93)
- UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (NeurIPS 2025 · similarity 0.91)
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.89)
- UltraDiff:Transferring High-Fidelity Priors to Compressed Latent Spaces for High-resolution Image Generation (NeurIPS 2026 · similarity 0.89)
- ResMaster: Mastering High-Resolution Image Generation via Structural and Fine-Grained Guidance (AAAI 2025 · similarity 0.88)
- Ultra-Resolution Adaptation with Ease (ICML 2025 · similarity 0.88)
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.
