🌟 Overview

RelightFormer revolutionizes image relighting by replacing traditional, computationally expensive inverse rendering with a feed-forward generative Transformer. By seamlessly injecting target lighting into spatial features and processing multiple views symmetrically, it delivers highly photorealistic results. Trained on the newly introduced, large-scale open-source Laval-Objaverse Dataset (LOD), RelightFormer achieves state-of-the-art quality and remarkable generalization across diverse scenes.

✨ Key Features

🏹 Feed-Forward Architecture: No iterative optimization required, enabling rapid generation.
🌟 Multi-View Consistency: Coherent and physically plausible relighting across all viewpoints.
⚡ Performant Inference: Highly optimized and expeditious execution on modern GPUs.
🎨 Competitive Quality: State-of-the-art, photorealistic relighting results.

🚀 Quick Start

You can easily load and run the model using the diffsynth library in our GitHub Repository. We provide two revisions: main (RelightFormer) and post (RelightFormer-Post, fine-tuned for enhanced quality).

from diffsynth import RelightFormerPipeline

# Use revision='main' for RelightFormer, or 'post' for RelightFormer-Post
pipe = RelightFormerPipeline.from_pretrained(
    "vLAR/RelightFormer", 
    revision="main"
)

# Example inference (adjust inputs according to your specific pipeline API)
# output = pipe(image=..., lighting=..., ...)

💡 For full inference scripts, multi-GPU evaluation, and training code, please visit the Official GitHub Repository.

📦 Dataset

This model is trained on the Laval-Objaverse Dataset (LOD), comprising 90,545 high-quality 3D assets and 39,008 diverse illumination conditions.

🤗 Browse/Download the Dataset: vLAR/LavalObjaverseDataset
📖 Rendering Instructions: See the RENDERING_INSTRUCTION.md in the GitHub repo.

🏋️ Training Details

RelightFormer is fine-tuned from the Wan 2.1 base model. The training pipeline consists of two stages:

Main Training: Trained on the full LOD dataset to learn multi-view relighting priors.
Post-Training: A secondary fine-tuning stage to further enhance photorealism and consistency.

Detailed training configurations, hardware requirements (e.g., 4× H200 GPUs), and scripts are available in the GitHub Repository.

📜 License

This model, its code, and associated datasets are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

🙏 Acknowledgements

This work was supported in part by the National Natural Science Foundation of China, the Research Grants Council of Hong Kong, the Otto Poon Charitable Foundation Smart Cities Research Institute, the Research Center for Unmanned Autonomous Systems, and the PolyU Kunpeng & Ascend Technology Innovation Incubation Center, The Hong Kong Polytechnic University.

Downloads last month: -; Downloads are not tracked for this model. How to track