TransLight: Image-Guided Customized Lighting Control with Generative Decoupling
Zongming Li, Lianghui Zhu, Haocheng Shen, Longjin Ran, Wenyu Liu, Xinggang Wang
TL;DR
TransLight is a novel framework that enables high-fidelity, customizable transfer of complex light effects between images by disentangling content and lighting through generative decoupling, advancing illumination editing capabilities.
Contribution
The paper introduces Generative Decoupling and IC-Light, enabling precise separation and transfer of light effects, which improves flexibility and fidelity in illumination stylization.
Findings
First method to transfer light effects across disparate images
Achieves more natural and customizable lighting control
Demonstrates superior performance over existing techniques
Abstract
Most existing illumination-editing approaches fail to simultaneously provide customized control of light effects and preserve content integrity. This makes them less effective for practical lighting stylization requirements, especially in the challenging task of transferring complex light effects from a reference image to a user-specified target image. To address this problem, we propose TransLight, a novel framework that enables high-fidelity and high-freedom transfer of light effects. Extracting the light effect from the reference image is the most critical and challenging step in our method. The difficulty lies in the complex geometric structure features embedded in light effects that are highly coupled with content in real-world scenarios. To achieve this, we first present Generative Decoupling, where two fine-tuned diffusion models are used to accurately separate image content and…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
- The formulation of the “light effect transfer” task—distinct from conventional relighting or style transfer—is novel and well-motivated. The generative decoupling framework using two diffusion models is an innovative approach to address content-light entanglement, which has not been effectively resolved in prior work. - The pipeline is well-engineered, with thoughtful design choices such as DINOv2-based filtering and multi-source relighting in training. Ablation studies confirm the effectivene
- Inconsistent training supervision for decoupling: The light removal model is trained with both synthesized (IS) and relighted images (IR), while the light extraction model uses only IS. This inconsistency is not justified and may lead to a domain gap in model robustness. A clearer rationale or ablation would strengthen the method section. - Threshold γ lacks theoretical justification: The cosine similarity threshold γ=0.98 is crucial in triplet filtering. However, the choice seems empirical an
1. The problem modeling is good. Lighting effects are important for real photo but less discussed in relighting research. 2. The data processing looks satisfying, and I encourage authors to release it. 3. The model design looks reasonable and foreseeably easy to reproduce (if with proper data)
1. From the results, it seems that rotational/directional illumination is less discussed. But I can understand that the main purpose is to achieve those lighting effects. 2. The model seems relatively weak in changing the illumination direction in its input images – many results display the inability in changing light direction. 3. Most results show very strong artistic lighting expressions but in real photos they are relatively rare, so the scope may be a bit narrow. But I understand that this
1) The method is simple and is an interesting decomposition into content vs. light effects 2) Scalable data pipeline for building triplets without manual labels. 3) Practical for creative workflows like image harmonization, where exact physical relighting is unnecessary. 4) The lighting removal is perhaps the most interesting part of the pipeline that can suppress strong lighting phenomena like glares, specular effects, etc.
1) One of the main claimed contributions of the paper is that it is“First to transfer light effects with geometric structure across disparate images,” which is not correct. Prior work already transfers lighting in a physically grounded way: Latent Intrinsics (cited) and its generative adaptation -- LumiNet: Latent Intrinsics meets diffusion models for indoor scene relighting (CVPR 2025; not cited) transfer lighting codes from a source to a target while handling large geometric mismatches in comp
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsImage Enhancement Techniques · Color Science and Applications · Building Energy and Comfort Optimization