🎨 LucidFlux:
Caption-Free Universal Image Restoration with a Large-Scale Diffusion Transformer

🌍 Website | 📘 Technical Report | 🧩 Models

---

---

News & Updates

---

Let us know if this works!

👥 Authors

Song Fei¹*, Tian Ye¹*‡, Lei Zhu^1,2†

¹The Hong Kong University of Science and Technology (Guangzhou)
²The Hong Kong University of Science and Technology

*Equal Contribution, ‡Project Leader, †Corresponding Author

---

🌟 What is LucidFlux?

LucidFlux is a framework designed to perform high-fidelity image restoration across a wide range of degradations without requiring textual captions. By combining a Flux-based DiT backbone with Light-weight Condition Module and SigLIP semantic alignment, LucidFlux enables caption-free guidance while preserving structural and semantic consistency, achieving superior restoration quality.

📊 Performance Benchmarks

📈 Quantitative Results

Benchmark	Metric	ResShift	StableSR	SinSR	SeeSR	DreamClear	SUPIR	LucidFlux (Ours)
RealSR	CLIP-IQA+ ↑	0.5005	0.4408	0.5416	0.6731	0.5331	0.5640	0.7074
	Q-Align ↑	3.1045	2.5087	3.3615	3.6073	3.0044	3.4682	3.7555
	MUSIQ ↑	49.50	39.98	57.95	67.57	49.48	55.68	70.20
	MANIQA ↑	0.2976	0.2356	0.3753	0.5087	0.3092	0.3426	0.5437
	NIMA ↑	4.7026	4.3639	4.8282	4.8957	4.4948	4.6401	5.1072
	CLIP-IQA ↑	0.5283	0.3521	0.6601	0.6993	0.5390	0.4857	0.6783
	NIQE ↓	9.0674	6.8733	6.4682	5.4594	5.2873	5.2819	4.2893
RealLQ250	CLIP-IQA+ ↑	0.5529	0.5804	0.6054	0.7034	0.6810	0.6532	0.7406
	Q-Align ↑	3.6318	3.5586	3.7451	4.1423	4.0640	4.1347	4.3935
	MUSIQ ↑	59.50	57.25	65.45	70.38	67.08	65.81	73.01
	MANIQA ↑	0.3397	0.2937	0.4230	0.4895	0.4400	0.3826	0.5589
	NIMA ↑	5.0624	5.0538	5.2397	5.3146	5.2200	5.0806	5.4836
	CLIP-IQA ↑	0.6129	0.5160	0.7166	0.7063	0.6950	0.5767	0.7122
	NIQE ↓	6.6326	4.6236	5.4425	4.4383	3.8700	3.6591	3.6742

---

🎭 Gallery & Examples

🎨 LucidFlux Gallery

---

🔍 Comparison with Open-Source Methods

LQ	SinSR	SeeSR	SUPIR	DreamClear	Ours

Show more examples

---

💼 Comparison with Commercial Models

LQ	HYPIR	Topaz	SeeDream 4.0	Gemini-NanoBanana	GPT-4o	Ours

Show more examples

---

🏗️ Model Architecture

Caption-Free Universal Image Restoration with a Large-Scale Diffusion Transformer

Our unified framework consists of four critical components in the training workflow:

🔤 Scaling Up Real-world High-Quality Data for Universal Image Restoration

🎨 Two Parallel Light-weight Condition Module Branches for Low-Quality Image Conditioning

🎯 Timestep and Layer-Adaptive Condition Injection

🔄 Semantic Priors from Siglip for Caption-Free Semantic Alignment

🚀 Quick Start

🔧 Installation

# Clone the repository
git clone https://github.com/W2GenAI-Lab/LucidFlux.git
cd LucidFlux

# Create conda environment
conda create -n lucidflux python=3.9
conda activate lucidflux

# Install dependencies
pip install -r requirements.txt

Inference

Prepare models in 2 steps, then run a single command.

1. Login to Hugging Face (required for gated FLUX.1-dev). Skip if already logged-in.

python -m tools.hf_login --token "$HF_TOKEN"

2. Download required weights to fixed paths and export env vars

# FLUX.1-dev (flow+ae), SwinIR prior, T5, CLIP, SigLIP and LucidFlux checkpoint to ./weights
python -m tools.download_weights --dest weights

# Exports FLUX_DEV_FLOW/FLUX_DEV_AE to your shell
source weights/env.sh

Run inference (uses fixed relative paths):

bash inference.sh

You can also obtain results of LucidFlux on RealSR and RealLQ250 from Hugging Face: LucidFlux.

🪪 License

The provided code and pre-trained weights are licensed under the FLUX.1 [dev].

🙏 Acknowledgments

• This code is based on FLUX. Some code are brought from DreamClear, x-flux. We thank the authors for their awesome work.

• 🏛️ Thanks to our affiliated institutions for their support.

• 🤝 Special thanks to the open-source community for inspiration.

---

📬 Contact

For any questions or inquiries, please reach out to us:

• Song Fei: [email protected]
• Tian Ye: [email protected]

🧑‍🤝‍🧑 WeChat Group

点击展开二维码（WeChat Group QR Code）