Research map / Image generation / Editing and style transfer
Consistent Feature Transport for Image Relighting
Bohan Zhang, huanweiliang huanweiliang, Yuhan He, Hongteng Xu, Xiaochao Qu, Luoqi Liu, Dixin Luo, Ting Liu
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 CFT training (L2 and L3) to Flux-Kontext (ablation) | CFT · Portrait relighting test set (446 pairs) | FID: 20.3817 -> 18.7338 | Improved |
| add CFT training to Qwen-Image-Edit (ablation) | CFT · Portrait relighting test set (446 pairs) | FID: 30.2147 -> 25.5529 | Improved |
| add CFT to Flux-Kontext (ablation) | CFT · Portrait relighting test set (446 pairs) | PSNR: 22.9192 -> 23.5105 | Improved |
| add only noise-to-source reconstruction loss L2 to L1 (ablation) | CFT · Portrait relighting ablation | FID: 20.3817 -> 21.6054 | Hurt |
| add transport loss L3 to L1 (ablation) | CFT · Portrait relighting ablation | PSNR: 22.9192 -> 23.3778 | Improved |
| add L2 on top of L1+L3 (ablation) | CFT · Portrait relighting ablation | PSNR: 23.3778 -> 23.5105 | Improved |
| use pairs sharing illumination transformation as L3 target instead of original pair (ablation) | CFT · Portrait relighting ablation | PSNR: 22.8031 -> 23.5105 | Improved |
| use random pairs instead of illumination-matched pairs for L3 (ablation) | CFT · Portrait relighting ablation | PSNR: 23.5105 -> 22.7958 | Hurt |
| increase alpha from 0.1 to 0.3 or 0.6 (ablation) | CFT · Portrait relighting | SSIM/PSNR/LPIPS/FID | Hurt |
| apply CFT in zero-shot transfer to real Multi-Illumination data | Flux-Kontext · Multi-Illumination Dataset | PSNR: 14.4246 -> 15.0590 | Improved |
| apply CFT to style transfer | Qwen-Image-Edit-2509 · Style transfer | PSNR: 12.9693 -> 13.3707 | Improved |
| apply CFT to style transfer | Qwen-Image-Edit-2509 · Style transfer | FID: 78.7016 -> 81.5195 | Hurt |
Closest papers by meaning
- UniLumos: Fast and Unified Image and Video Relighting with Physics-Plausible Feedback (NeurIPS 2025 · similarity 0.89)
- SynthLight: Portrait Relighting with Diffusion Model by Learning to Re-render Synthetic Faces (CVPR 2025 · similarity 0.89)
- LumiNet: Latent Intrinsics Meets Diffusion Models for Indoor Scene Relighting (CVPR 2025 · similarity 0.88)
- UniRelight: Learning Joint Decomposition and Synthesis for Video Relighting (NeurIPS 2025 · similarity 0.88)
- Lux Post Facto: Learning Portrait Performance Relighting with Conditional Video Diffusion and a Hybrid Dataset (CVPR 2025 · similarity 0.88)
- DreamLight: Towards Harmonious and Consistent Image Relighting (NeurIPS 2025 · similarity 0.88)
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.
