GFPGAN: Generative Facial Prior GAN

repository·master·Indexed 12 days ago

A generative facial prior model for real-world blind face restoration. It restores high-quality facial details from degraded images using pre-trained models like StyleGAN2 and ArcFace. Supports multiple model versions (V1, V1.2, V1.3) to balance naturalness and sharpness, and provides a CLI for inference via inference_gfpgan.py.

Tokens: 2.7K
Snippets: 6
Records: 10
Agent score: 96%

What's inside GFPGAN

Compare GFPGAN Model Versions

master

GFPGAN provides different model versions depending on your needs for naturalness vs. sharpness:

Version	Model Name	Description
V1.3	`GFPGANv1.3.pth`	More natural restoration results. Better for very low-quality or high-quality inputs. Can have repeated restorations. Weakness: Not very sharp, slight identity change.
V1.2	`GFPGANCleanv1-NoCE-C2.pth`	No colorization; no CUDA extensions required. Sharper output with beauty makeup. Weakness: Some outputs can be unnatural.
V1	`GFPGANv1.pth`	The original paper model, includes colorization.

Note: V1.3 is not always better than V1.2; choose based on whether you prioritize naturalness or sharpness.

Compare GFP-GAN model versions (V1.2 vs V1.3)
master
When choosing between GFP-GAN model versions, consider your specific input quality and desired output characteristics. V1.3 is not universally better than V1.2; the optimal choice depends on your use case.
V1.3
Best for:
- Very low-quality inputs.
- Relatively high-quality inputs.
- Achieving natural-looking outputs.
Trade-offs:
- Outputs may not be as sharp.
- May cause a slight change in facial identity.
- Can occasionally result in repeated (twice) restorations.
V1.2
Best for:
- Producing sharper outputs.
- Applying a 'beauty makeup' effect.
Trade-offs:
- Some outputs may appear unnatural.
https://github.com/tencentarc/gfpgan/blob/master/Comparisons.md

Run Quick Inference with GFPGAN

master

To perform inference, first download the pre-trained model GFPGANv1.pth into the experiments/pretrained_models directory.

Download Model

wget https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth -P experiments/pretrained_models

Run Inference

Depending on your installation method, use the following commands:

If using JIT installation: You must prefix your command with BASICSR_JIT=True.

BASICSR_JIT=True python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1

If you compiled extensions during installation:

python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1

For Aligned Images: Add the --aligned flag to either method if your input images are already aligned.

# Example for JIT
BASICSR_JIT=True python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1 --aligned

wget https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth -P experiments/pretrained_models

# JIT Inference
BASICSR_JIT=True python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1

# Compiled Inference
python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1

Download and organize pre-trained models
master
To use GFP-GAN, you must download the required pre-trained models and place them in the experiments/pretrained_models/ directory. The following files are necessary for the model to function:
- StyleGAN2 model: StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth
- FFHQ component locations: FFHQ_eye_mouth_landmarks_512.pth
- ArcFace model: arcface_resnet18.pth
```
# Example directory structure
gfpgan/
└── experiments/
    └── pretrained_models/
        ├── StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth
        ├── FFHQ_eye_mouth_landmarks_512.pth
        └── arcface_resnet18.pth
```
https://github.com/tencentarc/gfpgan/blob/master/experiments/pretrained_models/README.md
Train GFPGAN
master
To train GFPGAN, follow these procedures:
1. Dataset Preparation: Use the FFHQ dataset.
2. Download Pre-trained Data: Place the following in experiments/pretrained_models:
  StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth
  FFHQ_eye_mouth_landmarks_512.pth
  arcface_resnet18.pth
3. Configure: Modify options/train_gfpgan_v1.yml.
4. Run Training: Use torch.distributed.launch to start training.
Tip: Using higher quality faces and performing pre-processing (like beauty makeup) can improve restoration quality.
```
python -m torch.distributed.launch --nproc_per_node=4 --master_port=22021 gfpgan/train.py -opt options/train_gfpgan_v1.yml --launcher pytorch
```
https://github.com/tencentarc/gfpgan/blob/master/README.md
Install the original GFPGAN model (Paper Version)
master
To use the original model described in the paper, you must compile or load customized PyTorch C++ extensions (StyleGAN2). There are two installation methods depending on whether you want to compile extensions upfront or load them just-in-time (JIT).
Prerequisites
1. Clone the repository:
```
git clone https://github.com/xinntao/GFPGAN.git
cd GFPGAN
```
Option 1: Just-In-Time (JIT) Loading
Best for simple inference tasks with fewer installation issues. You must set the BASICSR_JIT=True environment variable before running your commands.
```
pip install basicsr
pip install facexlib
pip install -r requirements.txt
python setup.py develop
```
Option 2: Compile Extensions During Installation
Recommended for frequent training or inference. This method compiles CUDA extensions in BasicSR during the pip installation process and may take several minutes.
```
BASICSR_EXT=True pip install basicsr -vvv
pip install facexlib
pip install -r requirements.txt
python setup.py develop
```
```
# Option 1: JIT
pip install basicsr
pip install facexlib
pip install -r requirements.txt
python setup.py develop

# Option 2: Compile during install
BASICSR_EXT=True pip install basicsr -vvv
pip install facexlib
pip install -r requirements.txt
python setup.py develop
```
https://github.com/tencentarc/gfpgan/blob/master/PaperModel.md
Where to place model weights
master
To use the pre-trained models for inference, you must place the downloaded weight files into the gfpgan/weights/ directory of the repository.
https://github.com/tencentarc/gfpgan/blob/master/gfpgan/weights/README.md
Install GFPGAN
master
To install the clean version of GFPGAN (which does not require customized CUDA extensions and can run on Windows or CPU), follow these steps:
1. Clone the repository:
  git clone https://github.com/TencentARC/GFPGAN.git cd GFPGAN
2. Install dependencies: You will need Python >= 3.7 and PyTorch >= 1.7. Use the following commands to install the required packages:
  pip install basicsr pip install facexlib pip install -r requirements.txt python setup.py develop
3. Optional: Background Enhancement: If you want to enhance non-face regions (the background) using Real-ESRGAN, install the realesrgan package:
  pip install realesrgan
```
git clone https://github.com/TencentARC/GFPGAN.git
cd GFPGAN

# Install dependencies
pip install basicsr
pip install facexlib
pip install -r requirements.txt
python setup.py develop

# Optional: for background enhancement
pip install realesrgan
```
https://github.com/tencentarc/gfpgan/blob/master/README.md

Reference: inference_gfpgan.py CLI flags

master

The inference_gfpgan.py script accepts the following command-line arguments:

Flag	Description
`-h`	Show this help
`-i input`	Input image or folder. Default: `inputs/whole_imgs`
`-o output`	Output folder. Default: `results`
`-v version`	GFPGAN model version. Options: `1`
`-s upscale`	The final upsampling scale of the image. Default: `2`
`-bg_upsampler`	Background upsampler. Default: `realesrgan`
`-bg_tile`	Tile size for background sampler, `0` for no tile during testing. Default: `400`
`-suffix`	Suffix of the restored faces
`-only_center_face`	Only restore the center face
`-aligned`	Input are aligned faces
`-ext`	Image extension. Options: `auto`

Usage: python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2 [options]...

  -h                   show this help
  -i input             Input image or folder. Default: inputs/whole_imgs
  -o output            Output folder. Default: results
  -v version           GFPGAN model version. Option: 1 | 1.2 | 1.3. Default: 1.3
  -s upscale           The final upsampling scale of the image. Default: 2
  -bg_upsampler        background upsampler. Default: realesrgan
  -bg_tile             Tile size for background sampler, 0 for no tile during testing. Default: 400
  -suffix              Suffix of the restored faces
  -only_center_face    Only restore the center face
  -aligned             Input are aligned faces
  -ext                 Image extension. Options: auto | jpg | png, auto means using the same extension as inputs. Default: auto

Inference CLI Arguments
master
The inference_gfpgan.py script accepts the following arguments for controlling the restoration process:
- --input: Path to the input images (e.g., inputs/whole_imgs).
- --output: Path where the restored images will be saved (e.g., results).
- --version: Specifies the model version (e.g., 1).
- --aligned: A flag used when the input images are already face-aligned.
https://github.com/tencentarc/gfpgan/blob/master/PaperModel.md

GFPGAN: Generative Facial Prior GAN

What's inside GFPGAN

V1.3

V1.2

Download Model

Run Inference

Prerequisites

Option 1: Just-In-Time (JIT) Loading

Option 2: Compile Extensions During Installation