Last updated: 3 September 2026
AI face restoration is a niche full of confidently stated numbers that trace back to nothing. Fabricated accuracy percentages, invented download counts, and screenshots of results no tool actually produced circulate freely because almost nobody checks. This page sets out the rules we hold ourselves to so that you can check us.
1. Our Principles
We test what we can, and say when we haven't
Installation guides are written against a real install. Where we describe a command, an error message, or a settings panel, we have run it. Where we are summarising someone else's reported result rather than our own, the sentence says so explicitly.
Every number needs a source
We do not publish benchmark figures, download counts, citation counts, or accuracy percentages unless we can point to where they come from. If a number cannot be attributed, it does not go on the page — even when a rounder, more impressive figure would read better.
We describe our own tool accurately
The free enhancer on this site is a classical Canvas-based image filter, not a neural network, and every page that mentions it says so. We will not describe our own software as something it is not in order to rank for a term.
Advertising never influences content
This site is funded by Google AdSense. Advertisers have no input into what we publish, no advance sight of it, and no ability to have a comparison changed. We do not accept payment for a favourable review or placement in a comparison table.
We are independent of the model's authors
GFPGAN is built by the ARC Lab at Tencent PCG. We have no relationship with them, receive nothing from them, and speak for nobody but ourselves. Criticism of the model appears here when it is warranted.
2. What We Accept as a Source
We rank sources in tiers. A claim is only as strong as the best tier supporting it, and claims that reach only the bottom tier are not published at all.
| Tier | What it covers | How we use it |
|---|---|---|
| Primary | The CVPR 2021 paper and its supplementary material; the TencentARC/GFPGAN repository, its README, its release notes, and its source code. | Preferred for any claim about how the model works or what it was trained to do. |
| First-hand testing | Our own local installs across Windows, macOS and Linux; hosted runs on Replicate and Hugging Face; AUTOMATIC1111, Forge and ComfyUI integrations. | Preferred for setup steps, error messages, timings, and practical workflow advice. |
| Secondary | Peer-reviewed papers that evaluate or extend GFPGAN; maintainer comments in GitHub issues; official documentation for adjacent tools. | Used for context, and always linked so you can check it. |
| Not used | Marketing copy from commercial restoration apps; AI-generated summaries of other blogs; unattributed statistics circulating on content-farm sites. | Excluded. These are the main vector for the fake benchmark numbers that circulate in this niche. |
3. How We Test
When a guide describes running GFPGAN, the underlying testing follows the same shape:
- Install from the reference repository at a pinned commit, in a clean virtual environment.
- Run against a fixed set of our own test images — scanned prints, compressed screenshots, low-light phone photos, and modern portraits — so results are comparable across models.
- Record the exact command, the model checkpoint, the hardware, and the wall-clock time.
- Repeat anything surprising before it is written down.
- Note where output quality degrades, not only where it succeeds.
Where a guide reports timings, the hardware is stated alongside them. A number without hardware context is meaningless, and we treat its absence elsewhere as a red flag.
4. Use of AI in Our Writing
We use AI tools for drafting assistance, editing, and code formatting. We do not publish unreviewed AI output. Every guide is checked line by line by a human against the sources above before it goes live, and factual claims introduced by a model are verified or cut. AI is not used to generate example results, testimonials, benchmark figures, or before-and-after images.
5. Corrections
We would rather be corrected than be wrong quietly. If you find an error:
- Email hello@gfpgaan.com or use the contact form.
- Point at the specific sentence — a URL and a quote is ideal.
- We aim to respond within five working days.
Substantive factual corrections are made in the page itself and noted with an updated modification date. We do not silently delete pages to make an error disappear. Where a correction changes the conclusion of a comparison, we say so in the text.
6. Review and Update Schedule
Installation and setup guides are re-checked when a dependency or the upstream repository changes in a way that breaks the documented steps. Comparison articles are reviewed when a new model version ships. Every page carries a publication date, and updated pages carry a modification date, so you can judge for yourself whether the advice is still current.
7. Advertising and Funding
This site carries advertising served by Google AdSense. That is our only revenue source at present. We run no popunder, pop-up, vignette, interstitial, or auto-redirect advertising, and we do not place ads inside content in a way that could be mistaken for editorial recommendation. If we ever add affiliate links, they will be disclosed on the page that carries them and listed here. See our Privacy Policy and Cookie Policy for the data implications.
8. Who Writes This
Authorship for every article is shown on the article itself and links to a profile listing that author's background and their other work. See About Us for who we are and why this site exists.