AI image model · Tencent ARC

GFPGAN Explained: Face Restoration, Versions and License

GFPGAN is an open-source model from Tencent ARC that restores blurry, low-quality faces in a single pass; this page covers how it works, which version to use, how to run it and what its license allows. Need to sharpen product or marketing images instead? Supavisual upscales any image 2x or 4x for 1 credit.

  • Supavisual: 3 free trial credits
  • Upscale 2x or 4x for 1 credit
  • Supavisual has no face-restoration model
Where the GFPGAN paper was published
CVPR 2021
GitHub stars on the GFPGAN repo, November 2026
37.7k
Newest GFPGAN model weights
v1.4
Per upscale on Supavisual
1 credit

01

What is GFPGAN?

GFPGAN is a face restoration model from Tencent ARC, the Applied Research Center of Tencent PCG. The name comes from the paper "GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior" by Xintao Wang, Yu Li, Honglun Zhang and Ying Shan, posted in January 2021 and published at CVPR 2021.

"Blind" restoration means the model doesn't know how a photo was damaged: blur, noise, compression and low resolution all arrive mixed together. GFPGAN handles that by borrowing a generative facial prior, the knowledge stored inside a pretrained face GAN such as StyleGAN2. Channel-split spatial feature transform layers feed that prior into the restoration network, which balances realism against fidelity to the input face.

The result: GFPGAN restores facial detail and enhances color in a single forward pass, while GAN-inversion methods needed slow optimization for every image. The code and pretrained weights are free on GitHub.

02

GFPGAN face restoration features

What the official GFPGAN repository and its inference script offer.

Finds and fixes faces

GFPGAN detects each face in a photo, restores it, and pastes it back. You can restore only the center face, or feed it faces that are already cropped and aligned.

Background upscaling

Non-face regions are enlarged with Real-ESRGAN by default, so the whole photo is upscaled, not just the faces. The default final scale is 2x.

Output scale and tiling

One setting picks the final upscale of the whole photo, 2x by default, and another tiles the background pass so large images fit in memory.

Several models

GFPGAN ships V1, V1.2, V1.3 and v1.4 weights, plus inference code for RestoreFormer, another face restoration model.

Runs without custom CUDA

The clean GFPGAN version needs no compiled CUDA extensions. An NVIDIA GPU with CUDA and Linux are both listed as optional.

Training code included

Training scripts use the FFHQ face dataset and a pretrained 512 px StyleGAN2, so you can train or adapt GFPGAN to your needs.

03

GFPGAN v1.4, V1.3, V1.2 and V1: which version to use

How GFPGAN's authors describe each model, plus the setup V1 needs. They say V1.3 is not always better than V1.2, so test on your inputs.

VersionWhat it isStrengthsWeaknesses
v1.4Successor to V1.3Slightly more detail and better identity than V1.3None listed separately
V1.3Based on V1.2More natural results, better on very low-quality and on fairly high-quality inputs, can be run twiceNot very sharp, slight change in identity
V1.2Clean model, no colorization, trained on more dataSharper output, a beauty-makeup lookSome outputs look unnatural
V1The model from the paper, with colorizationMatches the published resultsNeeds the original setup with CUDA extensions

04

GFPGAN specs

What Tencent ARC has published for GFPGAN as of November 2026.

Maker
Tencent ARC (Applied Research Center, Tencent PCG)
Paper
GFP-GAN, arXiv January 2021, CVPR 2021 (pages 9168–9178)
Task
Blind face restoration with color enhancement
Prior
Pretrained face GAN (StyleGAN2), via channel-split spatial feature transform layers
Latest weights
v1.4
Latest package release
v1.3.8, September 16, 2022; last commit April 2024
Requirements
Python 3.7+, PyTorch 1.7+; NVIDIA GPU and CUDA optional
License
Apache 2.0, with third-party parts under their own licenses
Alternative: Supavisual upscale
Any image, 2x or 4x, 1 credit, no install

05

How to use GFPGAN online or on your own computer

The quickest way to try GFPGAN online is the Replicate demo linked from the official repo. As of November 2026 Replicate lists it at about $0.0029 a run, typically finished within 4 seconds, with more than 116 million runs, and lets you pick v1.2, v1.3, v1.4 or RestoreFormer, with v1.4 as the default. The repo also links demos on Hugging Face Spaces and Colab; on November 15, 2026 the main Space showed a runtime error, so it may not load.

To run GFPGAN locally, clone the repository, install basicsr, facexlib and the requirements, then realesrgan if you want the background upscaled. Download a weights file such as GFPGANv1.4.pth and run inference_gfpgan.py on an image or folder, choosing the version with -v and the final scale with -s. Results land in a results folder, with the restored face crops saved separately from the full image.

06

Is GFPGAN free for commercial use?

GFPGAN itself is released under the Apache License 2.0, which allows commercial use. That grant comes with an exception for listed third-party components, and two of those raise non-commercial questions. The StyleGAN2 code is modified from stylegan2-pytorch, and the license text also reproduces NVIDIA's StyleGAN2 license, which limits use to research or evaluation. Code derived from DFDNet is under Creative Commons Attribution-NonCommercial-ShareAlike 4.0.

Which of those parts your setup actually runs depends on the version and the path you use, so read the license file and check with counsel before selling a GFPGAN-based service. Restoring your own family photos is a different matter from building a product on it.

Remember, too, that the model reconstructs faces rather than recovering lost pixels. Its authors list a slight identity change as a V1.3 weakness, so a restored face is a plausible guess, not proof of what someone looked like.

07

GFPGAN limits and how to work around them

GFPGAN is strong on small, damaged faces. These are the gaps to plan for.

Do

  • Compare v1.4 with V1.3 and V1.2 on your own photos
  • Install Real-ESRGAN so the background is upscaled too
  • Try aligned crops if face detection misses a face
  • Check the third-party licenses before commercial use

Don't

  • Expect identical identity; restored faces can drift
  • Use it for products, text or scenes; GFPGAN targets faces
  • Present a restored face as a forensic record
  • Rely on the Hugging Face Space being online

08

A GFPGAN alternative for product and marketing images

Supavisual does not run GFPGAN and has no face-restoration model. If your job is reviving an old family portrait, use GFPGAN or a dedicated restorer. Supavisual is for a different problem: product and marketing images that are too small, too plain or the wrong shape.

In the editor, Upscale enlarges the whole image 2x or 4x, prompt edit and Retouch fix details you describe in words, background removal gives a transparent cutout, and Expand outpaints to a new aspect ratio. Smart resize turns one picture into social formats, and batch applies one edit to up to 10,000 images. Every action costs 1 credit.

Hobbyist is $19 a month for 800 credits, about 2.4 cents an action, with a $9 first month. Unlimited is $99 per seat per month. Both include a commercial license, and new accounts get 3 trial credits with watermarked previews.

09

GFPGAN vs Supavisual upscale: which one fits your photo?

Two different tools for two different jobs, as of November 2026.

FeatureGFPGANSupavisual
Built forRestoring damaged or low-quality facesProduct and marketing images
Face-specific restorationYesNo
UpscalingFaces restored, background via Real-ESRGAN, 2x by defaultWhole image, 2x or 4x
Edit with wordsNoYes, prompt edit and Retouch
Background removal, expand, resizeNoYes, 1 credit each
SetupPython install, or a hosted demoWeb app, no install
CostFree code; about $0.0029 a run on Replicate1 credit, about 2.4 cents on Hobbyist
LicenseApache 2.0 plus third-party termsCommercial license on paid plans

Read about other image models: GPT Image 2.6, Nano Banana 2.1, Seedream 5.5, MiniMax-H3-Image.

10

More models and Supavisual tools

Other model explainers and the tools that turn one image into a full set of visuals.

11

Frequently asked questions

What is GFPGAN?

GFPGAN is an open-source face restoration model from Tencent ARC, published at CVPR 2021. It uses the knowledge inside a pretrained face GAN, StyleGAN2, to rebuild detail and color in blurry or damaged faces in a single pass. It ships on GitHub with code, pretrained weights up to v1.4 and training scripts.

Is there a free GFPGAN online demo?

The official repo links demos on Replicate, Hugging Face Spaces and Colab. As of November 2026 Replicate charges about $0.0029 a run and typically finishes within 4 seconds. On November 15, 2026 the Hugging Face Space showed a runtime error, so Replicate or Colab are the safer picks. For product photos rather than faces, Supavisual's upscaler is another option.

What is the difference between GFPGAN v1.4 and V1.3?

GFPGAN's authors describe v1.4 as giving slightly more detail and better identity than V1.3. V1.3 aimed for more natural results and works on very low-quality inputs, but it is not very sharp and can slightly change identity. They also note V1.3 is not always better than V1.2, so pick by purpose and test on your own photos.

Can GFPGAN restore old photos?

It restores the faces in them. GFPGAN detects each face, rebuilds detail and color, and pastes it back, while Real-ESRGAN can upscale the rest of the picture. The repo lists no repair of scratches or tears outside faces. Treat the result as a reconstruction, since identity can shift slightly.

Can I use GFPGAN commercially?

The GFPGAN code is Apache 2.0, which permits commercial use, but its license file excludes third-party components. Its StyleGAN2 code comes with NVIDIA's StyleGAN2 license text, limited to research or evaluation, and DFDNet-derived code is CC BY-NC-SA 4.0. Check which parts your setup uses before building a paid product on it.

Does Supavisual use GFPGAN?

No. Supavisual does not run GFPGAN and has no face-restoration model. It is a separate image tool for products and marketing: it upscales whole images 2x or 4x, edits them from a written prompt, removes backgrounds and resizes them, each for 1 credit. New accounts get 3 free trial credits with watermarked previews.

What do I need to run GFPGAN locally?

Python 3.7 or newer and PyTorch 1.7 or newer. An NVIDIA GPU with CUDA speeds it up but is optional, and the clean version needs no custom CUDA extensions. Install basicsr, facexlib and the requirements, add realesrgan for backgrounds, download a weights file such as GFPGANv1.4.pth, then run the inference script.

Need sharper product images, not restored faces?

Read up on GFPGAN, then try Supavisual on your own product photo: upscale 2x or 4x, edit with a prompt or remove the background for 1 credit each. Start with 3 free trial credits.

Upscale an image

More ai image models: see all.

GFPGAN: Face Restoration Model, v1.4, License & Online Demo