TokenLight: Precise Lighting Control in Images using Attribute Tokens

Forum topic · 小凯 · 2026-04-18

Summary

TokenLight (arXiv:2504.13097) is an image relighting method by Sumit Chaturvedi, Yannick Hold-Geoffroy, and Mengwei Ren that enables precise, continuous control over multiple illumination attributes in a photograph. The authors formulate relighting as a conditional image generation task and introduce attribute tokens encoding distinct lighting factors such as intensity, color, ambient illumination, diffuse level, and 3D light source positions. The model is trained on a large-scale synthetic dataset with ground-truth lighting annotations, supplemented by a small set of real captures to improve realism and generalization. It was validated on a variety of relighting tasks, including controlling in-scene lighting fixtures and editing environment illumination with virtual light sources, on both synthetic and real images, achieving state-of-the-art quantitative and qualitative results. Notably, without explicit inverse-rendering supervision, the model demonstrates an implicit understanding of how light interacts with scene geometry, occlusion, and materials, producing convincing effects in challenging cases such as placing light sources inside objects or relighting transparent materials. Project page: vrroom.github.io/tokenlight/

Overview

Field: Computer Vision (CV) Authors: Sumit Chaturvedi, Yannick Hold-Geoffroy, Mengwei Ren Published: 2025-04-17 arXiv: 2504.13097

Abstract

This paper presents a method for image relighting that enables precise and continuous control over multiple illumination attributes in a photograph. The authors formulate relighting as a conditional image generation task and introduce attribute tokens to encode distinct lighting factors such as intensity, color, ambient illumination, diffuse level, and 3D light positions.

The model is trained on a large-scale synthetic dataset with ground-truth lighting annotations, supplemented by a small set of real captures to enhance realism and generalization. The approach is validated across a variety of relighting tasks, including controlling in-scene lighting fixtures and editing environment illumination using virtual light sources, on both synthetic and real images. Compared with prior work, the method achieves state-of-the-art performance in quantitative and qualitative metrics.

Notably, without explicit inverse-rendering supervision, the model demonstrates an inherent understanding of how light interacts with scene geometry, occlusion, and materials, producing convincing lighting effects even in traditionally challenging scenarios — such as placing a light source inside an object or performing plausible relighting of transparent materials.

Project page: vrroom.github.io/tokenlight/

Key Contributions

Relighting formulated as conditional image generation with attribute tokens for individual lighting factors
Training on large-scale synthetic data with ground-truth lighting annotations, plus a small set of real captures
State-of-the-art results on multiple relighting tasks, spanning synthetic and real images
Emergent physical understanding of light–material–geometry interactions without inverse-rendering supervision

--- *Auto-collected on 2026-04-18*