GFPGAN Tutorial: Restore Old Faces with AI — Complete Step-by-Step Guide

GFPGAN Tutorial: Restore Old Faces with AI — Complete Guide

This hands-on tutorial walks you through installing GFPGAN, restoring old face photos, enhancing video face restoration, and deploying with Docker. Each step includes exact commands you can copy and paste.

Prerequisites

• Python 3.7+ / pip
• NVIDIA GPU 2GB+ VRAM (CPU slower)
• Docker optional for container deployment

Step 1: Install

git clone https://github.com/TencentARC/GFPGAN.git && cd GFPGAN
pip install basicsr facexlib && pip install -r requirements.txt
python setup.py develop
wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth -P experiments/pretrained_models/

Step 2: Restore a Photo

python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2

Step 3: Cropped Face Mode

python inference_gfpgan.py -i inputs/cropped_faces -o results -v 1.4 -s 2 --aligned

Step 4: Full Photo (Faces + Background)

pip install realesrgan
python inference_gfpgan.py -i inputs/ -o results/ -v 1.3 -s 2 --bg_upsampler realesrgan --bg_tile 400

Step 5: Docker Deploy

docker pull wpafbo79/gfpgan:latest
docker run -d --name gfpgan --gpus all -v ./inputs:/inputs -v ./outputs:/outputs wpafbo79/gfpgan:latest
docker exec gfpgan python inference_gfpgan.py -i /inputs -o /outputs -v 1.3 -s 2

Step 6: Python API

import cv2
from gfpgan import GFPGANer
restorer = GFPGANer(model_path="GFPGANv1.3.pth", upscale=2, arch="clean", channel_multiplier=2)
img = cv2.imread("photo.jpg", cv2.IMREAD_COLOR)
_, _, restored = restorer.enhance(img, has_aligned=False, paste_back=True)
cv2.imwrite("restored.jpg", restored)

Model Guide

Troubleshooting

OOM Error

Use --bg_tile 200 or -s 1 for no upscale.

Face Not Detected

Crop faces manually and use --aligned mode.

Too Smooth

Use --weight 0.3 or try -v 1.2 model.

Slow on CPU

Use -s 1, cropped faces, or the Clean model variant.

API Wrapper

from flask import Flask, request, send_file
import cv2, tempfile
from gfpgan import GFPGANer
app = Flask(__name__)
r = GFPGANer(model_path="GFPGANv1.3.pth", upscale=2)
@app.route("/restore", methods=["POST"])
def restore():
    f = request.files["image"]; t = tempfile.NamedTemporaryFile(suffix=".jpg", delete=False)
    f.save(t.name); img = cv2.imread(t.name); _, _, out = r.enhance(img, paste_back=True)
    o = tempfile.NamedTemporaryFile(suffix=".jpg", delete=False)
    cv2.imwrite(o.name, out)
    return send_file(o.name, mimetype="image/jpeg")
app.run(host="0.0.0.0", port=5000)

Conclusion

GFPGAN takes ~10 minutes to set up. For best results pair with Real-ESRGAN for complete photo restoration.