Research map / Image generation / Image restoration
Every Pixel Has its Moments: Ultra-High-Resolution Unpaired Image-to-Image Translation via Dense Normalization
Ming-Yang Ho, Che-Ming Wu, Min-Sheng Wu, Yufeng Jane Tseng
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 |
|---|---|---|---|
| estimate pixel-level statistics (DN) instead of image-level (TIN) statistics | CUT inference · real2paint | FID: 249.612 -> 237.202 | Improved |
| estimate pixel-level statistics (DN) instead of patch-level kernelized statistics (KIN) | CUT inference · ANHIR (breast) | FID: 91.443 -> 68.616 | Improved |
| replace patch-wise IN with DN at inference | CUT · ANHIR (breast) | FID: 64.202 (patch-wise IN) vs 68.616 | No clear effect |
| use DN instead of TIN for domain differentiation | CUT · real2paint | domain differentiation accuracy: 0.556 -> 0.986 | Improved |
| use DN instead of patch-wise IN, TIN or KIN in human evaluation | CUT · summer2autumn (experts, best quality) | preferred (%): 82.86 (DN) vs 11.43 (KIN) | Improved |
| reformulate bilinear interpolation as fast interpolation (ablation) | DN · 4,302x3,024 image | runtime (s): 192.50 -> 8.28 | Improved |
| precompute interpolation matrices (ablation) | DN · 4,302x3,024 image | runtime (s): 8.28 -> 5.51 | Improved |
| add prefetching parallelism to run caching and inference in one pass (ablation) | DN · 4,302x3,024 image | runtime (s): 5.51 -> 4.35 | Improved |
| increase interpolating granularity from 1 pixel to 512 pixels (ablation) | DN · qualitative ablation | tiling artifacts | Hurt |
| run DN single-pass versus KIN two-stage | DN · 4,302x3,024 image | GPU VRAM (mb): 3145 (KIN) -> 4157 | Hurt |
Closest papers by meaning
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · similarity 0.88)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.87)
- L2P: Unlocking Latent Potential for Pixel Generation (NeurIPS 2026 · similarity 0.85)
- Is One GPU Enough? Pushing Image Generation at Higher-Resolutions with Foundation Models. (NeurIPS 2024 · similarity 0.85)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.85)
- UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (NeurIPS 2025 · similarity 0.84)
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.
