Research map / Image generation / Text-to-image diffusion
DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Training-free progressive high-resolution synthesis that upsamples SDXL outputs, re-noises and denoises them, and injects DWT low-frequency structure from the interpolated image plus sharpening to reach 2K–4K resolutions.
Problem
Pretrained text-to-image diffusion models such as SDXL produce repetitive objects and distorted shapes when sampling far above their training resolution, and existing fixes need retraining or are slow or fragile.
Approach
Use the generated low-resolution image, interpolated up, as structural guidance: re-noise it and denoise progressively, replacing the low-frequency DWT approximation coefficients of the estimated clean image with those of the interpolated image during early denoising steps, plus a sharpening step to counter blur.
Key results
- At 4096x4096, DiffuseHigh reaches FID_r 56.09 and FID_p 38.93 versus 62.82 and 48.97 for the next best training-free baseline DemoFusion.
- DiffuseHigh takes 258 sec per 4K image on a single A100 versus 728 sec for DemoFusion at 4096x4096.
- Against SR baselines, DiffuseHigh gives better FID_p and KID_p at 2048x4096 and 4096x4096, while SR models score slightly better on FID_r and KID_r.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffuseHigh instead of DemoFusion | SDXL high-resolution image synthesis · LAION-5B, 2048 × 2048 | FID_r: 57.16 -> 56.90 | Helped |
| Use DiffuseHigh instead of DemoFusion | SDXL high-resolution image synthesis · LAION-5B, 2048 × 4096 | FID_p: 49.08 -> 33.73 | Helped |
| Use DiffuseHigh instead of DemoFusion | SDXL high-resolution image synthesis · LAION-5B, 4096 × 4096 | FID_r: 62.82 -> 56.09 | Helped |
| Use DiffuseHigh instead of DemoFusion | SDXL high-resolution image synthesis · LAION-5B, 2048 × 2048 | CLIP score: 31.95 -> 31.26 | Hurt |
| Add DWT-based structural guidance (ablation) | DiffuseHigh · LAION-5B, 4096 × 4096 | FID_r: 56.24 -> 56.09 | Helped |
| Add sharpening (ablation) | DiffuseHigh · LAION-5B, 4096 × 4096 | FID_r: 56.39 -> 56.09 | Helped |
| Add both DWT structural guidance and sharpening (ablation) | DiffuseHigh · LAION-5B, 4096 × 4096 | FID_r: 56.49 -> 56.09 | Helped |
| Apply sharpening before noising and denoising (ablation) | SDXL on Gaussian-blurred inputs · Toy denoising experiment, Gaussian Blur | FID_r: 3.28 -> 2.70 | Helped |
| Apply sharpening before noising and denoising (ablation) | SDXL on downsampled-then-upsampled inputs · Toy denoising experiment, DownUp | FID_r: 3.07 -> 2.61 | Helped |
| Set structural guidance steps to δ=5 rather than δ=3 (ablation) | DiffuseHigh · LAION-5B, 4096 × 4096 | FID_r: 56.35 -> 56.09 | Helped |
| Increase structural guidance steps from δ=5 to δ=7 (ablation) | DiffuseHigh · LAION-5B, 4096 × 4096 | FID_r: 56.09 -> 56.18 | Hurt |
| Increase structural guidance steps from δ=5 to δ=10 (ablation) | DiffuseHigh · LAION-5B, 4096 × 4096 | FID_r: 56.09 -> 56.49 | Hurt |
Surprising findings
- Simply sharpening the interpolated image noticeably improves quality, which the authors attribute to diffusion priors favouring blurry samples.
- Applying structural guidance only in early denoising steps (5 of 15) worked best, and larger guidance counts reduced high-frequency detail.
Limitations
- Structural defects in the low-resolution image are propagated to the high-resolution output.
- Sharpness factor and guidance steps need some tuning; large sharpening causes artifacts.
- Evaluation relies on FID/KID/CLIP on a 1K caption subset without human study.
Experimental setup
Datasets: LAION-5B
Benchmarks: LAION-5B 1K caption subset
Metrics: FID_r KID_r FID_p KID_p CLIP score inference time
Baselines: ScaleCrafter FouriScale HiDiffusion DemoFusion SDXL+SD-Upscaler SDXL+BSRGAN
Builds on: SDXL MultiDiffusion Make-a-Cheap-Scaling EDM
Closest papers by meaning
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.92)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.91)
- ResMaster: Mastering High-Resolution Image Generation via Structural and Fine-Grained Guidance (AAAI 2025 · similarity 0.90)
- ScaleDiff: Higher-Resolution Image Synthesis via Efficient and Model-Agnostic Diffusion (NeurIPS 2025 · similarity 0.90)
- RepLDM: Reprogramming Pretrained Latent Diffusion Models for High-Quality, High-Efficiency, High-Resolution Image Generation (NeurIPS 2025 · similarity 0.90)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 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.