Research map / Image generation / Text-to-image diffusion
Creatively Upscaling Images with Global-Regional Priors
Yurui Qian, Qi Cai, Yingwei Pan, Ting Yao, Tao Mei
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 Global Structure Prior (ablation) | C-Upscale · LAION-5B synthetic 2048x2048 | MUSIQ: 73.11 -> 74.02 | Improved |
| add Regional Attention Prior (ablation) | C-Upscale · LAION-5B synthetic 2048x2048 | CLIP score: 0.8608 -> 0.8687 | Improved |
| add Regional Semantic Prior (ablation) | C-Upscale · LAION-5B synthetic 2048x2048 | MUSIQ: 72.80 -> 74.02 | Improved |
| add Regional Semantic Prior (ablation) | C-Upscale · LAION-5B synthetic 2048x2048 | FID: 67.62 -> 66.72 | Improved |
| add Global Structure Prior (ablation) | C-Upscale · LAION-5B synthetic 2048x2048 | LPIPS: 0.2793 -> 0.2766 | Improved |
| increase GSP step size s from 0.2 to 1.0 (ablation) | C-Upscale GSP · qualitative, SDXL | visual artifacts: s = 0, 0.2, 1.0 | Hurt |
| increase GSP step size from 0 to 0.2 (cosine schedule) (ablation) | C-Upscale GSP · qualitative, SDXL | object repetition: s = 0, 0.2, and 1.0 | Improved |
| use linear decreasing instead of cosine scheduler in GSP (ablation) | C-Upscale GSP · qualitative, SDXL | visual artifacts | Hurt |
| use linear increasing scheduler in GSP (ablation) | C-Upscale GSP · qualitative, SDXL | sharpness | Hurt |
| increase region overlap from 512 to 768 (ablation) | C-Upscale · SDXL 4096x4096 | inference time: {0,512,768} | Hurt |
| replace progressive upscaling and 50-step noise denoising with diffuse-then-denoise | DemoFusion-style upscaling · LAION-5B synthetic 4096x4096 | inference time: 14min -> 4min | Improved |
| use Hyper-SDXL few-step model and vLLM for LLaVA | C-Upscale · 1024 to 4096 upscaling | inference time: ~ 4 minutes to ~ 30 seconds | 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.
- DemoFusion — DemoFusion: Democratising High-Resolution Image Generation With No $$$ (CVPR 2024) · better on 10 of 10 shared results
- ScaleCrafter — ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024) · better on 1 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
- The Power of Context: How Multimodality Improves Image Super-Resolution (CVPR 2025 · similarity 0.89)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.88)
- Scale Where It Matters: Training-Free Localized Scaling for Diffusion Models (NeurIPS 2026 · similarity 0.88)
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.88)
- Local Conditional Controlling for Text-to-Image Diffusion Models (AAAI 2025 · similarity 0.88)
- Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis (CVPR 2025 · similarity 0.87)
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.
