Research map / 3D vision / Face and shape
DiFaReli++: Diffusion Face Relighting With Consistent Cast Shadows
Puntawat Ponglertnapakorn, Nontawat Tritrong, Supasorn Suwajanakorn
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 Modulator network for spatial shading-reference conditioning (ablation) | DiFaReli · Multi-PIE ablation | DSSIM: 0.0749 -> 0.0670 | Improved |
| use shading reference instead of light encoding as non-spatial vector (ablation) | DiFaReli · Multi-PIE ablation | DSSIM: 0.0885 -> 0.0670 | Improved |
| add s, cam and ArcFace xi non-spatial conditions (ablation) | DiFaReli · Multi-PIE ablation | DSSIM: 0.0713 -> 0.0670 | Improved |
| add DECA s and cam conditions (ablation) | DiFaReli · Multi-PIE ablation | MSE: 0.0063 -> 0.0077 | Hurt |
| add DECA s and cam conditions (ablation) | DiFaReli · Multi-PIE ablation | DSSIM: 0.0674 -> 0.0670 | Improved |
| add ArcFace xi condition (ablation) | DiFaReli · Multi-PIE ablation | DSSIM: 0.0686 -> 0.0670 | Improved |
| remove zero-conv block from ControlNet (CN-mod) (ablation) | ControlNet + cross-attention · Multi-PIE conditioning alternatives | DSSIM: 0.0983 -> 0.0697 | Improved |
| replace ControlNet spatial conditioning with DiFaReli++ Modulator (ablation) | cross-attention non-spatial · Multi-PIE conditioning alternatives | DSSIM: 0.0983 -> 0.0694 | Improved |
| replace cross-attention with DiFaReli++ non-spatial conditioning (ablation) | DiFaReli++ spatial · Multi-PIE conditioning alternatives | DSSIM: 0.0694 -> 0.0664 | Improved |
| add cast shadow map conditioning (DiFaReli++ over DiFaReli) | DiFaReli · Multi-PIE (same subject as target lighting) | DSSIM: 0.0711 -> 0.0604 | Improved |
| distill DiFaReli++ into single-shot network | DiFaReli++ · Multi-PIE (same subject as target lighting) | DSSIM: 0.0604 -> 0.0590 | Improved |
| distill into single-shot network instead of multi-step DPM-Solver/UniPC sampling | DiFaReli++ · Runtime | seconds: 0.07 seconds vs over one second at 10 steps; 13.56s at 100 steps | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiffusionRig — DiffusionRig: Learning Personalized Priors for Facial Appearance Editing (CVPR 2023) · better on 2 of 2 shared results
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
- DiFaReli: Diffusion Face Relighting (ICCV 2023 · similarity 0.99)
- SynthLight: Portrait Relighting with Diffusion Model by Learning to Re-render Synthetic Faces (CVPR 2025 · similarity 0.89)
- Lux Post Facto: Learning Portrait Performance Relighting with Conditional Video Diffusion and a Hybrid Dataset (CVPR 2025 · similarity 0.88)
- Neural Gaffer: Relighting Any Object via Diffusion (NeurIPS 2024 · similarity 0.86)
- IllumiNeRF: 3D Relighting Without Inverse Rendering (NeurIPS 2024 · similarity 0.86)
- BecomingLit: Relightable Gaussian Avatars with Hybrid Neural Shading (NeurIPS 2025 · similarity 0.86)
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.
