Who Developed GFPGAN? TencentARC, Downloads, and Running It on Replicate

GFPGAN is one of the most downloaded AI face restoration models ever built. But a lot of people who use it every day have never heard the name behind it — TencentARC.

This article covers everything you need to know:

Who exactly built GFPGAN
Where to find the official TencentARC/GFPGAN repository
How to download GFPGAN model weights
How GFPGAN works as a photo restoration and image enhancer
How to run TencentARC GFPGAN on Replicate in minutes

Whether you want to use the browser tool, run it in the cloud, or download the model weights for local use — this guide covers all your options.

Who Developed GFPGAN?

GFPGAN was developed by a team of researchers at Tencent ARC — the Applied Research Center of Tencent, one of the largest technology companies in the world. The core team behind the model was:

Xintao Wang — lead researcher, also known for work on Real-ESRGAN and BasicSR
Yu Li — researcher at Tencent ARC
Honglun Zhang — researcher at Tencent ARC
Ying Shan — senior researcher and project lead at Tencent ARC

The paper was titled “Towards Real-World Blind Face Restoration with Generative Facial Priors” and was accepted at CVPR 2021 — one of the most respected computer vision research conferences in the world. CVPR acceptance is highly competitive, which speaks to the quality and originality of the work.

The team published GFPGAN as open-source software under the MIT License. This means anyone — students, hobbyists, businesses, and developers — can download, use, and modify it for free.

Since its release in 2021, GFPGAN has:

Accumulated over 10 million downloads
Been cited in 400+ academic papers
Been integrated into Stable Diffusion WebUI, ComfyUI, and dozens of other tools
Spawned multiple follow-up models from other research groups

The lead researcher, Xintao Wang, has since continued to work on related projects including Real-ESRGAN (a general-purpose image upscaler) and CodeFormer contributions, making Tencent ARC one of the most influential labs in practical AI image restoration.

What Is the TencentARC/GFPGAN Repository?

The official home of GFPGAN is the TencentARC/GFPGAN repository on GitHub. You can find it by searching “TencentARC GFPGAN” on GitHub or going directly to the repository under the TencentARC organization.

The repository contains:

The full model source code written in Python (PyTorch)
Pre-trained model weight files for GFPGAN v1.2, v1.3, v1.4, and the unofficial community v1.5
Inference scripts to run restoration from the command line
A training script if you want to fine-tune the model on your own data
Detailed installation instructions for Windows, macOS, and Linux
A colab notebook for running GFPGAN in Google Colab without local setup

The repository is actively maintained and has thousands of GitHub stars, making it one of the most starred face AI projects on the platform.

GFPGAN Model Versions

The repository offers four main model versions, each with different trade-offs:

Version	Output Size	Best For	File Size
GFPGAN v1.2	512 px	Fast preview, mild damage	~332 MB
GFPGAN v1.3	512 px	Balanced — preserves more original character	~332 MB
GFPGAN v1.4	1024 px	Recommended default — high fidelity	~332 MB
GFPGAN v1.5	1024 px	Maximum detail recovery (community)	~332 MB

For most GFPGAN photo restoration tasks, v1.4 is the best starting point. It outputs at 1024 pixels with very high identity fidelity and handles the widest range of damage types well.

How to Download GFPGAN

You have three main options, depending on what you want to do.

Option 1 — Use It Free in Your Browser (No Download Needed)

The fastest way to use GFPGAN is right here on this site. No download, no Python, no setup. Just upload your photo and restore it in seconds. Everything runs in your browser using WebAssembly — your photo never leaves your device.

This is the best option for:

Restoring individual photos quickly
Trying GFPGAN before committing to a local setup
Anyone who does not want to deal with Python environments

Restore a photo free now →

Option 2 — Download from GitHub (Local Setup)

To download and run GFPGAN locally, follow these steps:

1. Clone the repository

git clone https://github.com/TencentARC/GFPGAN.git
cd GFPGAN

2. Install dependencies

pip install basicsr facexlib realesrgan
pip install -r requirements.txt
python setup.py develop

3. Download a model weight file

Go to the Releases section of the TencentARC/GFPGAN repository and download the .pth model file for the version you want. Place it in the experiments/pretrained_models/ folder.

4. Run inference

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

This restores all images in the inputs/whole_imgs folder and saves results to the results folder. The -s 2 flag upscales the background by 2× using Real-ESRGAN.

For a full walkthrough, see our GFPGAN installation guide.

Option 3 — Run on Replicate (Cloud, No Local Setup)

Replicate lets you run GFPGAN in the cloud — no Python, no GPU required. We cover this in detail in the section below.

GFPGAN as a Photo Restoration Tool

When most people say GFPGAN photo restoration, they mean using it to fix old or damaged portrait photos. This is genuinely one of the best use cases in all of AI image processing.

GFPGAN photo restoration works because the model was specifically trained on face images. It does not just sharpen pixels — it rebuilds face detail using a generative facial prior derived from StyleGAN2. This is what separates it from basic sharpening filters or general upscalers.

What GFPGAN restores in photos:

Old family prints — film grain, faded colors, scan artifacts, low dynamic range
Blurry portraits — motion blur, focus blur, camera shake
Compressed images — JPEG block artifacts from heavy lossy compression
Low-light shots — high-ISO grain and color noise from dark indoor photos
Pixelated crops — small face regions from group or wide-angle shots
Black and white scans — noise, low contrast, and soft detail in archival prints

The restoration happens in four stages: face detection → degradation encoding → generative prior injection → blending back into the original image. The entire pipeline runs in under two seconds in the browser and under one second on a local GPU.

Want to see actual results across all these photo types? Check out our GFPGAN Before & After Examples (2026) with real metrics.

GFPGAN as an Image Enhancer

GFPGAN is often called a GFPGAN image enhancer — and while that description is accurate, it is important to understand what kind of enhancement it does.

GFPGAN enhances faces specifically. It is not a general-purpose image enhancer like Real-ESRGAN or Topaz Photo AI. It will not sharpen a landscape, improve a product photo, or fix a blurry dog. It targets human facial regions only.

Within that scope, it is exceptional. Here is what makes GFPGAN stand out as a face image enhancer:

Natural results, not artificial ones. Basic sharpening makes edges pop but introduces halos and artifacts. GFPGAN reconstructs from real face knowledge — the result looks like it was always a good photo.

Identity-preserving enhancement. Other enhancement tools can make a face look sharper but also make it look like a different person. GFPGAN achieves 98.2% FaceNet cosine similarity — meaning the enhanced face looks like the same person, just clearer.

Works on severely degraded input. Most image enhancers fail on truly damaged photos. GFPGAN was designed specifically for blind face restoration — where the type and level of damage is unknown. This makes it far more useful in real-world photo restoration scenarios.

Pairs well with Real-ESRGAN. For full-image enhancement, the recommended workflow is: GFPGAN for the face regions first, then Real-ESRGAN for the background and full image. This two-pass approach gives you complete, polished results.

How to Run TencentARC GFPGAN on Replicate

Replicate is a cloud platform that lets you run AI models through a web interface or API — without installing anything on your computer. It is one of the most popular ways to run TencentARC GFPGAN without dealing with Python environments or GPU requirements.

Step 1 — Find GFPGAN on Replicate

Go to replicate.com and search for “GFPGAN” or “TencentARC”. You will find the official GFPGAN model listed under the TencentARC organization. The model page shows example outputs, input parameters, and the API documentation.

Step 2 — Create a Free Account

Replicate requires a free account to run models. Sign up at replicate.com — you get a small amount of free compute credit to get started.

Step 3 — Upload Your Photo and Set Parameters

On the GFPGAN model page, you will see an input panel on the left side. The key parameters are:

img — upload your photo here (JPG or PNG)
version — choose your model version (v1.4 recommended)
scale — upscaling factor for the background (2 is standard)
codeformer_fidelity — if the Replicate version includes CodeFormer blending, set this between 0 and 1 (higher = more faithful to original, lower = more AI reconstruction)

Step 4 — Click Run

Hit the Run button. Replicate queues your job and typically processes it within a few seconds to a few minutes depending on server load. The result appears on the same page as a downloadable image.

Step 5 — Download Your Result

Click the output image to download it in full resolution. Replicate also shows you the processing time and resource usage.

Running GFPGAN on Replicate via API

For developers who want to build GFPGAN into an application, Replicate provides a clean REST API. Here is the basic structure in Python:

import replicate

output = replicate.run(
    "tencentarc/gfpgan:9283608cc6b7be6b65a8e44983db012355f829a",
    input={
        "img": open("your-photo.jpg", "rb"),
        "version": "v1.4",
        "scale": 2
    }
)

print(output)

The response includes a URL to the restored image hosted on Replicate’s CDN. You can download it or pass it directly to another step in your pipeline.

Replicate Costs

Replicate charges by compute time. GFPGAN typically costs a fraction of a cent per image on Replicate’s hardware. For individual use, the free credits are usually enough for testing. For production use with high volume, costs add up — the browser tool or local setup would be more economical.

Which Option Is Right for You?

Here is a quick summary to help you choose:

Use Case	Best Option
Restore one or two photos quickly	Free browser tool on this site
Privacy-sensitive images	Free browser tool (zero upload)
Batch processing many photos	Local setup (GitHub download)
API integration in an app	Replicate API
Testing without any setup	Replicate web interface
Full control over pipeline	Local GitHub clone

For most people, the free browser tool is the fastest and most private option. No account. No download. No image ever leaves your device.

Try GFPGAN Free — Right Now

No Download. No Replicate Account. Just Upload and Restore.

Runs entirely in your browser. 100% private. Nothing leaves your device.

Restore a Photo Free →

What Is GFPGAN? Complete Guide — Full explanation of how GFPGAN works, what it restores, and its limitations.
How to Install GFPGAN on Windows, Mac & Linux — Step-by-step local setup guide with screenshots.
GFPGAN Face Restoration: Before & After Examples (2026) — Real results with PSNR/SSIM metrics across six photo types.
CodeFormer vs GFPGAN: Which Is Better? — Deep comparison of the two leading face restoration models.
Best AI Face Restoration Models: GFPGAN, CodeFormer, GPEN — Full three-way model comparison.

Frequently Asked Questions

Who made GFPGAN?

GFPGAN was developed by Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan at Tencent ARC (Applied Research Center). It was published at CVPR 2021. The official code is maintained in the TencentARC/GFPGAN repository on GitHub under the MIT License.

Where can I download GFPGAN?

You can download GFPGAN from the official TencentARC/GFPGAN GitHub repository. The model weight files (.pth) are available in the Releases section. Alternatively, use the free browser tool on this site — no download required — or run it via the Replicate cloud platform.

Is GFPGAN free to download and use?

Yes. GFPGAN is released under the MIT License, which means it is completely free to download, use, modify, and redistribute — including for commercial projects. There are no license fees or usage limits.

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

GFPGAN v1.4 outputs at 1024px (vs 512px for v1.3) and has higher identity fidelity. v1.3 sometimes preserves more of the original character in moderately damaged photos, while v1.4 produces more aggressive and detailed reconstruction. For most users, v1.4 is the recommended default.

How do I run GFPGAN on Replicate for free?

Sign up for a free account at replicate.com, find the TencentARC/GFPGAN model, upload your photo, and click Run. Replicate gives you free compute credits when you sign up, which is enough to test the model on several photos. Each image typically takes a few seconds to process.

Can I use the Replicate API to build an app with GFPGAN?

Yes. Replicate provides a REST API and official Python, JavaScript, and Node.js clients. You can call the TencentARC/GFPGAN model programmatically and integrate face restoration into your own application. Replicate charges by compute time, so costs scale with your usage volume.