Real-ESRGAN
Real-ESRGAN — practical blind super-resolution for real-world photos and video: restores and upscales degraded images with models trained on pure synthetic data, BSD-3-Clause
GFPGANは、TencentARCが公開しているオープンソースの画像プロジェクトです(GitHub: TencentARC/GFPGAN)。GitHubスターは37,703、フォークは6,279で、Apache-2.0ライセンスの下で公開されています。GitHubでの最終更新は2024年7月26日です。
Tencent ARC face restoration — revives old, blurry and low-quality faces with a StyleGAN2 prior, Apache-2.0
Download the source and move into the project folder.
Install BasicSR, facexlib and the requirements file, then install the package in development mode.
Grab a pretrained model such as GFPGANv1.3.pth into experiments/pretrained_models.
Point the inference script at a folder of images and collect the results.
Rebuilds faces from blurry, noisy or compressed photos using a pretrained StyleGAN2 prior.
Runs on Windows and CPU — no custom CUDA build step.
v1.3 for natural results, v1.4 for detail and identity, v1.2 for sharpness.
Pairs with Real-ESRGAN to upscale the non-face parts of the photo too.
The full training pipeline is in the repo if you want to improve it on your own data.
Clone the repo and install its Python dependencies.
git clone https://github.com/TencentARC/GFPGAN.git
cd GFPGAN
pip install basicsr facexlib
pip install -r requirements.txt
python setup.py developFetch the pretrained weights into the models folder.
wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth -P experiments/pretrained_modelsRun inference at 2x upscaling; results land in the results folder.
python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2Close neighbours on the Image shelf: Real-ESRGAN — practical blind super-resolution for real-world, and CodeFormer — NeurIPS 2022] Towards Robust Blind Face Restoration. The full shelf is the Image tools in the catalogue.
It restores low-quality faces — blurry, compressed, noisy or old photos — using the facial knowledge inside a pretrained StyleGAN2.
v1.3 for the most natural results, v1.4 for more detail and identity, v1.2 if you want sharper output. Try more than one on your photos.
No — the clean version runs on CPU and Windows without CUDA extensions, though a GPU is faster.
Yes — Apache 2.0.
Yes, with the Real-ESRGAN background upsampler (pip install realesrgan).
Real-ESRGAN — practical blind super-resolution for real-world photos and video: restores and upscales degraded images with models trained on pure synthetic data, BSD-3-Clause
[NeurIPS 2022] Towards Robust Blind Face Restoration with Codebook Lookup Transformer
Visual diagnostics for LoRA training — spot overfitting with face biometrics
Simple Flask web app that upscales low-res images with AI
PhotoCraft — a clean-room Photoshop reimplementation in pure Rust: layers, masks, adjustment layers, type, vectors, brushes and real PSD files, offline, MIT / Apache-2.0
SAM 2 — Meta's Segment Anything Model 2: promptable segmentation for images and video, Apache-2.0
LibreAIカタログに掲載された画像プロジェクトの中で最もスターの多いもの — GFPGANが合わなかった場合に検討したい、現実的な候補リストです。
Stable Diffusion web UI
The node-graph powerhouse for precise, professional AI image and video generation.
High performance self-hosted photo and video management solution
The original latent text-to-image diffusion model from CompVis — the CVPR 2022 reference implementation, Open RAIL-M
PhotoCraft — a clean-room Photoshop reimplementation in pure Rust: layers, masks, adjustment layers, type, vectors, brushes and real PSD files, offline, MIT / Apache-2.0
カタログの中でGFPGANとスター数が近い画像プロジェクト — 前後の棚もそのまま見て回れます。
Animated sprite editor & pixel art tool (Windows, macOS, Linux)
Real-ESRGAN — practical blind super-resolution for real-world photos and video: restores and upscales degraded images with models trained on pure synthetic data, BSD-3-Clause
AI-Powered Photos App 🌈💎✨
🤗 Diffusers: State-of-the-art diffusion models for image, video, and audio generation in PyTorch.
🆙 Upscayl - #1 Free and Open Source AI Image Upscaler for Linux, MacOS and Windows
ControlNet — the original conditional control for diffusion models: steer Stable Diffusion with edges, depth, pose and scribbles instead of prompts alone, Apache-2.0