mon3tr/FluxSR-private

FluxSR

FluxSR is a LoRA adapter for one-step real-world image super-resolution, based on FLUX.1-dev. Training, inference, and evaluation code is publicly available in the FluxSR project. The adapter is provided as pytorch_lora_weights.safetensors.

Files and use

• pytorch_lora_weights.safetensors: transformer LoRA adapter.
• training_config.json: training and inference settings.
• t5_embedding.pth, clip_embedding.pth: the fixed text conditions.
• prompts_10k.json: the 10,000 prompts used to generate training pairs.
• FLUX.1-dev-LICENSE.md: model license supplied with the base model.
• NOTICE: BFL attribution and FluxSR modification notices.

Use the FluxSR inference entry point described in the project's README, with this directory as the adapter path. The FLUX.1-dev base model and the included fixed T5/CLIP text embeddings are also needed. Download the base model from BFL after accepting its access conditions. The model package provides the LoRA weights, configuration, fixed text embeddings, and training prompts. Model binaries and JSON files are hosted on Hugging Face and are not tracked in the GitHub code repository.

From the FluxSR code directory:

python inference.py \
  --input /path/to/input.png --output results/example \
  --model /path/to/FLUX.1-dev --lora weights/fluxsr --seed 114

The input can also be a directory. Inference performs ×4 super-resolution and reads timestep, guidance, LoRA rank, and alpha from training_config.json. The fixed embeddings are loaded from the adapter directory. CPU offload (--offload sequential) and VAE tiling (--vae-tiling) are optional. The evaluation below uses full-image inference.

This model repository is public. Download the complete package:

hf download mon3tr/FluxSR-private --local-dir weights/fluxsr

The package contains the FluxSR adapter and supporting files. Download the FLUX.1-dev base model separately from Black Forest Labs.

Training configuration

Setting	Value
Base and teacher	FLUX.1-dev
Generated pairs	10,000 prompts, 1024 × 1024
Training steps	20,000
Teacher sampling	50 steps, guidance 3.5, base seed 10
Training batch	1 per GPU, 8 GPUs; global batch 8
LoRA rank / alpha	64 / 64
Learning rate / warmup	1e-5 / 500 steps
Student conditioning timestep / guidance	350 / 4.0
Losses	Spatial MSE FTD, LPIPS 1.0, TV-LPIPS 1.0, ADL 0.1
Training precision	BF16 frozen weights; FP32 trainable adapters
Training seed	114

ADL averages the cosine similarity between each image token and the mean image-token representation across all 57 transformer blocks.

RealSR ×4 evaluation

Evaluation of the supplied weights:

Model	PSNR ↑	SSIM ↑	LPIPS ↓	DISTS ↓	MUSIQ ↑	MANIQA ↑	TOPIQ-NR ↑	Q-Align ↑
FluxSR	25.15367	0.73374	0.2964075	0.1870468	68.1300	0.5086289	0.6266896	4.382539

The evaluation uses all 100 Canon/Nikon test pairs from RealSR (V3), scale 4, full-image inference with BF16 frozen base weights and FP32 LoRA parameters, timestep 350, guidance 4, and seed 114 with the sorted image index added per image. No color correction or tiled inference was used. PSNR and SSIM use the Y channel with border crop 0; the other metrics use the documented PyIQA defaults. Refer to the project README for the exact evaluation settings and separate Q-Align environment.

Intended use and limitations

Intended for non-commercial research on real-world image super-resolution under the accompanying model license. As a generative restoration model, FluxSR may synthesize details that are absent from the input; restored content should not be treated as factual evidence. Large images require more GPU memory. Changing precision, tiling, image preprocessing, or random seeds may change outputs and measured scores.

License and attribution

The adapter is a modification of FLUX.1-dev. The adapter and bundled fixed T5/CLIP text-embedding tensors are subject to the FLUX.1 [dev] Non-Commercial License. See NOTICE for the required BFL attribution and modification notice. Rights to the base model and derivatives are granted directly by BFL under that license. The Apache-2.0 license of the project code does not grant additional rights to these model assets. This adapter is not an official BFL product and is not endorsed by BFL.

Citation

@inproceedings{li2025one,
  title={One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation},
  author={Li, Jianze and Cao, Jiezhang and Guo, Yong and Li, Wenbo and Zhang, Yulun},
  booktitle={ICML},
  year={2025}
}