Research map / Image generation / Editing and style transfer
Close Imitation of Expert Retouching for Black-and-White Photography
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.
Builds a retouching-aware black-and-white decolorization network (DeColorful-Net) that uses hierarchical proxy-based metric learning and a multi-level bilateral grid to imitate three experts' BW retouching styles, trained on a new expert-retouched BW version of MIT-Adobe 5K.
Problem
Producing aesthetically pleasing black-and-white photographs usually requires expert retouching, and existing decolorization methods focus on structure/contrast preservation rather than artistic style, which varies by expert and subject.
Approach
Treats decolorization as aesthetic imitation: subject-style proxies learned with a hierarchical proxy-based loss in a deep metric learning space condition a bilateral-grid network (with Dual-AdaIN) that applies spatially varying, multi-level retouching selected by user-chosen style.
Key results
- Ours (L1 + LCE) reaches 27.50 PSNR, 0.947 SSIM and 0.057 LPIPS on the expert-retouched test set, above the best listed baseline (HDRNet 23.93 PSNR, MAXIM 0.184 LPIPS worse), with Table 1 showing the ablation gains.
- In a 80-participant MTurk user study, the DeColorful-Net variants obtain the top average preference score (4.58 total for the Ours column in Study 2) over commercial BW filters and the experts' camera BW photos.
- In Study 3, users selected the proxy-matched style for internet-photographer styles at a 73.94% average rate, indicating scalability to proxies built from unpaired internet BW photos.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use L1 + LCE instead of L1 loss alone (ablation) | DeColorful-Net · B&W Adobe 5K test | PSNR: 27.34 -> 27.50 | Helped |
| use L1 + LCE instead of L2 loss (ablation) | DeColorful-Net · B&W Adobe 5K test | PSNR: 27.16 -> 27.50 | Helped |
| use L1 instead of L2 while retaining LCE (ablation) | DeColorful-Net · B&W Adobe 5K test | PSNR: 27.21 -> 27.50 | Helped |
| use Proxy-Anchor loss instead of triplet loss (ablation) | DeColorful-Net · B&W Adobe 5K test | PSNR: 26.35 -> 27.50 | Helped |
| add hierarchical proxy-based loss instead of using Proxy-Anchor loss alone (ablation) | DeColorful-Net · B&W Adobe 5K test | PSNR: 26.00 -> 27.50 | Helped |
| use two-stage training with proxy generation before decolorization (ablation) | DeColorful-Net · B&W Adobe 5K test | PSNR: 25.14 -> 27.50 | Helped |
| add a hierarchical multi-level bilateral grid (ablation) | DeColorful-Net · B&W Adobe 5K test | PSNR: 26.79 -> 27.50 | Helped |
Surprising findings
- Experts' own camera BW photos scored poorly against the comparison methods, attributed to narrower spectral range of color DSLR sensors.
- General image enhancement methods (e.g., HDRNet, StarEnhancer) that perform well on color retouching performed worse on BW images than the proposed method.
Limitations
- Proxy bank limited to 4 subjects and 4 styles in the main dataset; internet-photo proxies use only unpaired images.
- Evaluation of aesthetics relies on user preference studies with a limited set of participants and not objective metrics for artistry.
Experimental setup
Datasets: MIT-Adobe 5K internet-downloaded BW photos from 8 professional photographers
Benchmarks: B&W Adobe 5K
Metrics: PSNR SSIM LPIPS user preference score user preference percentage
Baselines: HDRNet StarEnhancer NeurOp MAXIM RSFNet CSNorm PDecolor LeDecolor SPDecolor iPhone 14 Pro BW filter Galaxy S21 Ultra BW filter Instagram filter
Builds on: Proxy-Anchor loss HDRNet bilateral grid Dual-AdaIN hierarchical proxy-based loss MIT-Adobe 5K
Closest papers by meaning
- DiffRetouch: Using Diffusion to Retouch on the Shoulder of Experts (AAAI 2025 · similarity 0.85)
- Training-and-Prompt-Free General Painterly Harmonization via Zero-Shot Disentenglement on Style and Content References (AAAI 2025 · similarity 0.84)
- Zero-shot Image Editing with Reference Imitation (NeurIPS 2024 · similarity 0.84)
- The Photographer's Eye: Teaching Multimodal Large Language Models to See, and Critique Like Photographers (CVPR 2025 · similarity 0.84)
- Automatic Controllable Colorization via Imagination (CVPR 2024 · similarity 0.83)
- JarvisArt: Liberating Human Artistic Creativity via an Intelligent Photo Retouching Agent (NeurIPS 2025 · similarity 0.83)
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.