Research map / Image generation / Fast diffusion sampling
DiffusionLight-Turbo: Accelerated Light Probes for Free via Single-Pass Chrome Ball Inpainting
Worameth Chinchuthakun, Pakkapon Phongthawee, Amit Raj, Varun Jampani, Pramook Khungurn, 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 Exposure LoRA (ablation) | DiffusionLight · Laval Indoor | angular error (mirror): 7.053 -> 5.936 | Improved |
| add iterative inpainting (ablation) | DiffusionLight · Laval Indoor | angular error (mirror): 6.146 -> 5.936 | Improved |
| add iterative inpainting (ablation) | DiffusionLight · Poly Haven | angular error (matte): 3.560 -> 3.595 | No clear effect |
| add LoRA swapping instead of Turbo-SDEdit (ablation) | DiffusionLight-Turbo · Laval Indoor | time (mins): 1.229 -> 0.535 | Improved |
| add LoRA swapping instead of Turbo-Pred (ablation) | DiffusionLight-Turbo · Laval Indoor | angular error (mirror): 6.449 -> 6.114 | Improved |
| apply Turbo LoRA for speedup | DiffusionLight · Laval Indoor | time (mins) and angular error (mirror): 31.889 -> 0.535 mins; 5.936 -> 6.114 | Mixed |
| use DiffusionLight instead of StyleLight | StyleLight (reproduced) · Poly Haven | angular error (mirror): 7.122 -> 4.286 | Improved |
| use Turbo instead of full DiffusionLight | DiffusionLight · Laval Indoor (array of spheres) | angular error: 5.25 -> 7.35 | Hurt |
| increase balls N from 5 to 30 (k=1) (ablation) | DiffusionLight · Poly Haven | angular error (mirror): 4.380 -> 4.282 | Improved |
| use two iterations rather than one (N=30) (ablation) | DiffusionLight · Poly Haven | angular error (mirror): 4.282 -> 4.286 | No clear effect |
| train Exposure LoRA on timesteps 900-999 rather than 0-999 (ablation) | Exposure LoRA · validation mirror | angular error: 6.292 -> 5.464 | Improved |
| train Turbo LoRA on timesteps 500-999 rather than 0-999 (ablation) | Turbo LoRA · validation mirror | angular error: 5.891 -> 5.539 | Improved |
Closest papers by meaning
- DiffusionLight: Light Probes for Free by Painting a Chrome Ball (CVPR 2024 · similarity 0.96)
- Neural LightRig: Unlocking Accurate Object Normal and Material Estimation with Multi-Light Diffusion (CVPR 2025 · similarity 0.88)
- LuxDiT: Lighting Estimation with Video Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
- Neural Gaffer: Relighting Any Object via Diffusion (NeurIPS 2024 · similarity 0.87)
- Diffusion Renderer: Neural Inverse and Forward Rendering with Video Diffusion Models (CVPR 2025 · similarity 0.87)
- Do Image Editing Models Understand Lighting? (NeurIPS 2026 · 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.
