Research map / Image generation / Image restoration
DiffRAW: Leveraging Diffusion Model to Generate DSLR-Comparable Perceptual Quality sRGB from Smartphone RAW Images
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.
Proposes DiffRAW, a diffusion model for smartphone RAW-to-DSLR-quality sRGB conversion that conditions on RAW and a color-position preserving image, plus a Domain Transform Diffusion Method that cuts inference steps.
Problem
Converting smartphone RAW images to DSLR-quality sRGB is ill-posed and hampered by detail loss, color mapping instability, and spatial misalignment in RAW-sRGB training pairs.
Approach
Uses a diffusion model conditioned on the RAW image (for structure) and a color-position preserving condition built from degraded DSLR images (for color and alignment), with a Domain Transform Diffusion Method that starts inference from a degraded-domain image and transitions toward the target domain, reducing steps.
Key results
- On ZRR no-reference metrics, DiffRAW reaches MUSIQ-K 56.67 vs 48.52 for LiteISPNet and DSLR reference 56.62.
- On ZRR with aligned GT, DiffRAW achieves LPIPS 0.118 and FID 14.61 vs LiteISPNet's 0.133 and 15.30.
- DTDM with 100 inference steps reportedly matches or surpasses DDPM with 1500 steps in detail (qualitative comparison).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffRAW instead of the best listed baseline | RAW-to-sRGB mapping · ZRR | MUSIQ-K: 48.52 -> 56.67 | Helped |
| Use DiffRAW instead of the best listed baseline | RAW-to-sRGB mapping · ZRR | LPIPS (Original GT): 0.187 -> 0.145 | Helped |
| Use DiffRAW instead of the best listed baseline | RAW-to-sRGB mapping · ZRR | PSNR (Original GT): 21.55 -> 21.31 | Hurt |
| Add RAW image condition w (ablation) | DiffRAW · ZRR | structural detail | Helped |
| Add color-position preserving condition c (ablation) | DiffRAW · ZRR | color and spatial consistency | Helped |
| Replace DDPM diffusion and reverse process with DTDM (ablation) | DiffRAW · ZRR | detail quality | Helped |
| Reduce inference denoising and iteration steps (ablation) | DiffRAW · ZRR | no-reference metrics | Hurt |
| Reduce inference denoising and iteration steps (ablation) | DiffRAW · ZRR | full-reference metrics | Helped |
Surprising findings
- DiffRAW's no-reference scores (MUSIQ-K, CLIPIQA+) match or exceed the DSLR reference images themselves.
Limitations
- PSNR and SSIM are only comparable, not superior, to baselines.
- Evaluated on a single dataset (ZRR).
- Fewer inference steps trade off no-reference metrics against full-reference metrics.
Experimental setup
Datasets: ZRR (Zurich RAW to RGB)
Benchmarks: ZRR
Metrics: LPIPS FID MUSIQ CLIPIQA+ PSNR SSIM NIQE IL-NIQE
Baselines: PyNet MW-ISPNet LiteISPNet DDPM
Builds on: DDPM LiteISPNet InDI
Closest papers by meaning
- ISPDiffuser: Learning RAW-to-sRGB Mappings with Texture-Aware Diffusion Models and Histogram-Guided Color Consistency (AAAI 2025 · similarity 0.90)
- Color Alignment in Diffusion (CVPR 2025 · similarity 0.86)
- Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range Tone Mapping (CVPR 2024 · similarity 0.86)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · similarity 0.86)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.85)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.85)
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.