Research map / 3D vision / 3D generation

SyncLight: Single-Edit Multi-View Relighting

NeurIPS 2026 · 3D vision › 3D generation · approach: Diffusion-based novel view synthesis · 0 citations (OpenAlex)

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 SyncLight, a pose-free multi-view diffusion transformer (SDXL with cross-view token attention) trained with latent bridge matching to propagate a single-view parametric light edit (intensity and Lab color via a lightmap) across uncalibrated views in one step.

Problem

Consistent parametric control of light sources across multiple uncalibrated views of a static indoor scene, where per-view single-image relighting produces geometrically inconsistent lighting and 3D methods need poses and per-scene optimization.

Approach

Replace SDXL self-attention with multi-view attention that concatenates view tokens so views interact, train with latent bridge matching (one-step inference) on image pairs with a lightmap conditioning on the reference view, and let it generalize zero-shot to N>2 views without poses.

Key results

• On the Real split reference view, SyncLight reaches 30.34 dB PSNR vs 28.42 for LightLab* and 22.31 for GR3EN.
• On the second view of the Real split, SyncLight retains 95.8% of ground-truth VGGT correspondences vs 82.7% for GR3EN.
• SyncLight runs in about 1.58 s for two views, versus 76.4 s for GR3EN on the second view, and is preferred in 73.9%-81.7% of user study votes.

Surprising findings

• One-step inference at test time outperforms multi-step inference even though the model was trained with four timesteps, which the authors attribute to starting from input images rather than noise.
• Training with a single timestep (equivalent to L2 regression) drops PSNR by 4.82 dB on the Real split, so multi-timestep supervision matters even for one-step inference.

Limitations

• Relies on a circular marker for light conditioning, so windows and unconventional light sources are not handled.
• Assumes a static scene, synchronized captures, and that edited lights are visible in the reference view.
• Compute grows quadratically with the number of views due to full cross-view attention.

Experimental setup

Datasets: Infinigen BlenderKit SyncLight dataset (real captures) LuxRemix synthetic OLAT data RealEstate10K

Benchmarks: SyncLight test set LuxRemix-derived external test set RealEstate10K (qualitative)

Metrics: PSNR SSIM Delta E 00 LPIPS VGGTm inference time user preference

Baselines: ScribbleLight LightLab* Flux.2-dev GR3EN LumiNet LuxRemix

Builds on: Latent Bridge Matching MVDream Stable Diffusion XL LightLab LumiNet

Closest papers by meaning

• UniRelight: Learning Joint Decomposition and Synthesis for Video Relighting (NeurIPS 2025 · similarity 0.90)
• 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 0.88)
• ReCap: Better Gaussian Relighting with Cross-Environment Captures (CVPR 2025 · similarity 0.87)
• Comprehensive Relighting: Generalizable and Consistent Monocular Human Relighting and Harmonization (CVPR 2025 · similarity 0.87)
• Generative Multiview Relighting for 3D Reconstruction under Extreme Illumination Variation (CVPR 2025 · similarity 0.87)