TokenLight: Precise Lighting Control in Images using Attribute Tokens

2026-04-16 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

TokenLight introduces a novel image relighting method that provides precise and continuous control over various illumination attributes within photographs. This approach frames relighting as a conditional image generation task, utilizing attribute tokens to represent distinct lighting factors like intensity, color, ambient light, diffuse levels, and 3D light positions. The model is trained on a large synthetic dataset featuring ground-truth lighting annotations, augmented with a smaller collection of real-world captures to improve realism and generalization. TokenLight demonstrates state-of-the-art quantitative and qualitative performance across diverse relighting tasks, including manipulating in-scene fixtures and editing environmental illumination with virtual light sources on both synthetic and real images. The model implicitly understands light-scene interactions, producing convincing effects even for complex scenarios like lights inside objects or relighting transparent materials.

Key takeaway

For research scientists developing image synthesis or editing tools, TokenLight demonstrates a robust approach to granular lighting control. You should consider integrating attribute tokens into your conditional image generation models to achieve precise manipulation of illumination, especially when working with complex scenes or requiring high realism. This method offers a pathway to more intuitive and physically plausible relighting capabilities without explicit inverse rendering.

Key insights

Attribute tokens enable precise, continuous control over multiple illumination factors in image relighting.

Principles

• Relighting can be framed as conditional image generation.
• Synthetic data with real data improves realism.

Method

The method uses attribute tokens to encode distinct lighting factors, training on a large synthetic dataset with ground-truth annotations, supplemented by real captures for enhanced generalization.

In practice

• Control in-scene lighting fixtures.
• Edit environment illumination with virtual lights.

Topics

• TokenLight
• Image Relighting
• Attribute Tokens
• Conditional Image Generation
• Illumination Control

Best for: Research Scientist, AI Scientist, Computer Vision Engineer

Related on AIssential

• Unified Pix Token And Word Token Generative Language Model — Unifying pixel and word tokens in generative models enhances visual detail understanding and enables unsupervised image pretraining.
• LiTo: Surface Light Field Tokenization — LiTo unifies 3D geometry and view-dependent appearance modeling using surface light field tokenization.
• PIXLRelight: Controllable Relighting via Intrinsic Conditioning — PIXLRelight enables physically accurate, controllable single-image relighting by bridging PBR and learned synthesis via intrinsic conditioning.
• Build with Nano Banana 2, our best image generation and editing model — Nano Banana 2 enhances image generation with improved world knowledge, text rendering, and creative control for scalable visual applications.
• OmniLight: One Model to Rule All Lighting Conditions — OmniLight and DINOLight models excel at image restoration under adverse lighting, achieving top NTIRE 2026 Challenge rankings.
• LUCID: Learning Unified Control for Image Deflaring and Exposure Mastery in Nighttime Photography — LUCID unifies controllable flare mitigation and exposure adjustment for superior nighttime image restoration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.