GFPGAN

GFPGAN is Tencent ARC’s open-source face restoration model that reconstructs realistic faces from blurry, degraded photos.

Overview

GFPGAN is the open-source reference for AI face restoration. It uses generative facial priors to rebuild sharp, believable faces from old, blurry, or low-resolution photos — fixing eyes, skin, and hair detail that generic upscalers hallucinate badly. The project spawned countless wrappers and web UIs, runs on modest GPUs, and is integrated in tools like ComfyUI workflows and Upscayl. For family photo restoration, it remains a default pipeline component.

GFPGAN’s approach — using a pretrained face prior network to guide restoration — became the template for most face repair tools that followed. Users can balance fidelity against restoration strength, and the model family offers faster variants for weaker hardware.

Key Features

• State-of-the-art face restoration from degraded inputs
• Realistic detail synthesis for eyes, skin, and hair
• Open-source weights and code for free use
• Runs locally on consumer GPUs
• Wide integration in AI image toolchains
• Multiple model versions and quality/speed tradeoffs

Restoration hobbyists bundle GFPGAN with upscalers and colorizers into pipelines that revive entire photo archives for free. The model’s influence on the field is hard to overstate: it defined what modern face restoration outputs should look like.

Pricing

GFPGAN is completely free and open source under a permissive license for research and commercial use.

Who Is It For

Photo restorers, ML engineers, and hobbyists building photo enhancement pipelines at zero cost.

FAQ

Is GFPGAN free to use?

Yes, entirely free and open source.

Does GFPGAN offer an API?

No official hosted API; it runs as local code, and third-party services wrap it.

How does GFPGAN compare to alternatives?

Consumer apps like Remini are more convenient; GFPGAN gives free, local, scriptable restoration with no upload of private photos.

Does GFPGAN restore non-face areas?

Backgrounds benefit modestly, but the model’s strength is faces; pair it with a general upscaler for scenes.

Is it safe for private photos?

Yes, running locally means photos never leave your machine.

Can I use GFPGAN commercially?

The license permits broad use; verify the current license text for your specific application.

How large are the model files?

Weights are compact enough for consumer GPUs and even some CPUs.

Is GFPGAN suitable for batch restoration?

Yes, the open code supports scripting batch jobs across entire photo collections.

Does it work on old photographs?

Yes, restoring aged and degraded portrait photos is its signature application.