Research map / Image generation / Super-resolution
Training-free, Perceptually Consistent Low-Resolution Previews with High-Resolution Image for Efficient Workflows of Diffusion Models
Wongi Jeong, Hoigi Seo, Se Young Chun
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 |
|---|---|---|---|
| select downsample matrix by minimum commutator norm instead of nearest downsampling (ablation) | Preview generation · FLUX.1-dev, 500 prompts | DreamSim: 10.39 -> 7.05 | Improved |
| select downsample matrix by minimum commutator norm instead of random (ablation) | Preview generation · FLUX.1-dev, 500 prompts | PSNR (dB): 20.851 -> 20.962 | Improved |
| select by argmin instead of argmax of commutator (ablation) | Preview generation · FLUX.1-dev, 500 prompts | DreamSim: 8.55 -> 7.05 | Improved |
| add commutator-zero guidance (ablation) | Preview generation · FLUX.1-dev, 500 prompts | DreamSim: 8.56 -> 7.05 | Improved |
| add commutator-zero guidance (ablation) | Preview generation · FLUX.1-dev, 500 prompts | PSNR (dB): 19.115 -> 20.962 | Improved |
| increase correction iterations m from 3 to 7 (ablation) | commutator-zero guidance · FLUX.1-dev PSNR sweep | PSNR (dB): 20.645 -> 21.389 at alpha=0.04 | Improved |
| increase step size alpha beyond optimum (ablation) | commutator-zero guidance · FLUX.1-dev PSNR sweep | PSNR (dB): 21.146 -> 20.543 (m=5, alpha 0.04 -> 0.06) | Hurt |
| increase m (more correction steps) (ablation) | commutator-zero guidance · compute cost | computation | Hurt |
| combine with TaylorSeer temporal-axis acceleration | Preview generation · FLUX.1-dev | DreamSim: 16.97 -> 7.79 vs FLUX.1-dev (10) at similar speedup | Improved |
| downsample at timestep tD with commutator-zero condition instead of naive downsampling | Preview generation · FLUX.1-dev | DreamSim: 9.20 -> 6.83 | Improved |
| downsample at timestep tD instead of generating directly at low resolution | Preview generation · FLUX.1-dev | DreamSim: 21.02 -> 6.83 | Improved |
| downsample latents instead of reducing NFE to 20 | Preview generation · SD3.5-L | PIQE: 29.39 ref; ours 31.55 vs naive 29.75 | Hurt |
Closest papers by meaning
- HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance (NeurIPS 2025 · similarity 0.89)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.89)
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.89)
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.89)
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.88)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · 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.
