Upload README.md with huggingface_hub

README.md

@@ -22,103 +22,3 @@ This repository hosts the released A<sup>2</sup>BFR **LoRA checkpoint** for infe
 - Code: [MediaX-SJTU/A2BFR](https://github.com/MediaX-SJTU/A2BFR)
 - Base model: [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev)
 - Checkpoint: `default.safetensors`
-
-## Model Details
-
-A<sup>2</sup>BFR injects the low-resolution face image as a visual condition and uses text prompts for attribute-aware restoration. The model is trained as a LoRA on top of FLUX.1-dev, so users must prepare the FLUX.1-dev base model separately and comply with its license and usage terms.
-
-Key features:
-
-- Blind face restoration from low-quality face inputs.
-- Text-guided attribute control during restoration.
-- LoRA checkpoint for efficient loading and sharing.
-- Inference code released in the GitHub repository.
-
-## Installation
-
-Clone the code repository and install dependencies:
-
-```bash
-git clone https://github.com/MediaX-SJTU/A2BFR.git
-cd A2BFR
-conda create -n a2bfr python=3.10 -y
-conda activate a2bfr
-pip install -r requirements.txt
-```
-
-Install a CUDA-compatible PyTorch build for your GPU environment if needed.
-
-## Download Weights
-
-```bash
-hf download thetrigger/A2BFR default.safetensors --local-dir checkpoints/A2BFR
-```
-
-## Inference
-
-Single image inference:
-
-```bash
-python infer.py \
-  --base-model /path/to/FLUX.1-dev \
-  --lora checkpoints/A2BFR/default.safetensors \
-  --condition-image /path/to/lq_face.png \
-  --output-dir outputs/demo \
-  --prompt "A photo of a human face" \
-  --guidance-scale 3.0 \
-  --seed 0 \
-  --concat
-```
-
-Batch inference:
-
-```bash
-python infer.py \
-  --base-model /path/to/FLUX.1-dev \
-  --lora checkpoints/A2BFR/default.safetensors \
-  --input-dir /path/to/lq_faces \
-  --output-dir outputs/batch \
-  --prompt "A high-quality, high-resolution, realistic, and extremely detailed image of a human face" \
-  --guidance-scale 3.0 \
-  --seed 0
-```
-
-Useful options:
-
-- `--prompt`: controls the target facial attributes and restoration semantics.
-- `--width` / `--height`: output resolution; default is `512 x 512`.
-- `--dtype`: defaults to `bfloat16`; use `float16` if needed.
-- `--preprocess deblurring`: creates a degraded condition image from a clean input. For already low-quality inputs, keep the default `--preprocess none`.
-- `--concat`: saves a side-by-side condition/result image.
-
-## Intended Use
-
-This model is intended for research on blind face restoration, text-guided image restoration, and controllable face image generation. It can be used to restore degraded face images and explore attribute-aware restoration behavior.
-
-## Limitations
-
-- The model requires FLUX.1-dev as the base model.
-- The released checkpoint is a LoRA and is not a standalone full model.
-- Results may vary with input quality, prompt wording, guidance scale, and random seed.
-- The model may fail on extreme degradations, heavy occlusions, non-face images, or images far outside the training distribution.
-- Generated facial attributes should not be treated as factual identity or biometric evidence.
-
-## Ethical Considerations
-
-Face restoration and attribute editing can be misused for impersonation, misleading media, or privacy-invasive applications. Users should obtain proper consent for face images, avoid deceptive use, and clearly disclose generated or restored content where appropriate.
-
-## Training
-
-Training code will be released.
-
-## Citation
-
-If A<sup>2</sup>BFR is useful for your work, please cite the project once the paper/citation is available.
-
-## License
-
-This checkpoint is released under the Apache 2.0 license. Users must also follow the license and acceptable use terms of the FLUX.1-dev base model.
-
-## Acknowledgement
-
-This project is based on [OminiControl](https://github.com/Yuanshi9815/OminiControl). We also leveraged [facer](https://github.com/FacePerceiver/facer)'s code in our project. Thanks for these awesome works.