Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DreamLight: Towards Harmonious and Consistent Image Relighting
Authors: Yong Liu, Wenpeng Xiao, Qianqian Wang, Junlin Chen, Shiyin Wang, Yitong Wang, Xinglong Wu, Yansong Tang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive comparisons and user study demonstrate that our Dream Light achieves remarkable relighting performance. ... Experiment results demonstrate that our Dream Light exhibits superior generalization and performance on the universal relighting of natural images. Related ablations also prove the rationality and effectiveness of the proposed designs. |
| Researcher Affiliation | Collaboration | Yong Liu1 , Wenpeng Xiao2 , Qianqian Wang2 , Junlin Chen2 , Shiyin Wang2 , Yitong Wang2 , Xinglong Wu2 , Yansong Tang1 1Tsinghua Shenzhen International Graduate School, Tsinghua University 2Byte Dance Inc. |
| Pseudocode | No | The paper describes the methodology in prose and uses figures (e.g., Figure 2 for the pipeline) to illustrate the process, but it does not contain a dedicated section or block labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | No | We will open source this relighting lora to benefit community. ... Codes and models need to be reviewed by our organization before they are made available. We will make them public as soon as possible. |
| Open Datasets | Yes | We design data generation pipeline to facilitate the training of our model. Our data has three sources. ... Secondly, we utilize available 3D assets [50] to render a number of consistent images with lighting of different color and directions. ... Reference [50] points to 'Objaverse: A universe of annotated 3d objects'. |
| Dataset Splits | No | Our data has three sources. ... Totally, the quantities of the three types of data are about 600k, 150k, and 300k, respectively. ... The evaluation benchmark contains 600 high-quality image pairs rendered by Arnold Renderer from real objects. While total training data quantities and an evaluation benchmark size are mentioned, specific training/validation splits or percentages for the overall dataset are not provided. |
| Hardware Specification | Yes | Our model is implemented in Py Torch [54] using 8 A100 at 512 512 resolutions. |
| Software Dependencies | Yes | Our model is implemented in Py Torch [54]... We leverage Stable Diffusion-v1.5 [1] as the base generative model and CLIP-H [55] as the encoder... we takes RMBG-1.4 as the segmentation model for extracting the region of subject. |
| Experiment Setup | Yes | Our model is implemented in Py Torch [54] using 8 A100 at 512 512 resolutions. The main model is trained end-to-end with the batch size of 512. The learning rate is set to 5e-5. ... The cutoff frequency of spectral filter is set to 5 in default. The number of light query is set to 4 for each direction. |