Why this exists
If you go searching for "AI background remover" or "free photo colorizer," you will find dozens of websites, all wrapping the same handful of open-source models. Most of them are paywalled past a free trial, plastered with intrusive ads, run aggressive email-capture flows, watermark the output, or quietly retain your uploads for "service improvement." A few of them are outright scams.
The underlying technology — U-2-Net, GFPGAN, Real-ESRGAN, AnimeGAN, and the others — is open-source. Anyone with a server, an afternoon, and some Python can run any of these models and offer them as a tool. The reason there are so many bad sites and so few good ones is that the financial incentives all push toward extracting more value from each visitor: subscriptions, watermarks, email gates, dark patterns. The friction is the business model.
This site exists as an experiment in not doing that. It is funded entirely by Google AdSense, and the ads are restricted to non-intrusive placements between content. There is no account, no email capture, no usage cap, no watermark, no upsell to a paid tier, no tracking analytics. If the ads cover the server bills, the site stays up. If they don't, eventually it won't.
Who runs it
The site is built and maintained by a single developer working in computer vision and machine learning. The infrastructure (a VPS for the web app, a separate machine for model inference, an HTTPS proxy in front) is privately managed; there is no third-party processing service involved. When you upload an image, it goes only to servers I personally administer.
That also means: when something breaks, it is one person fixing it. When a model produces an unexpected result, there is no support team — but if you email me about it, I'll read the email myself.
Why I write the guides
The other half of this site is the guides section: long-form articles explaining how each model actually works. Most "AI image editor" sites won't tell you what model they're using, much less how it works. I think that's silly. The technology is genuinely interesting, the trade-offs are worth understanding, and a user who knows what GFPGAN is doing under the hood will get better results than one who treats it as a black box.
The guides are written in plain language but assume you are a curious adult. They explain architectures, training data, and failure modes honestly — including the cases where the models fail in ways that matter ethically (face restoration on images of people of colour, for instance, has well-documented bias issues that come from skewed training data). These caveats appear in the guides because they should appear; not mentioning them would be dishonest.
The technology
Every tool on the site is powered by an open-source deep learning model. None of these models were trained by me — they are the work of researchers at universities and corporate labs, generously released under permissive licences. My contribution is the FastAPI server, the frontend, the worker pipeline, and the editorial decisions about how to present each one.
- Remove Background — U-2-Net (Qin et al., 2020). Salient object detection.
- Photo Colorizer — Colorful Image Colorization (Zhang, Isola, Efros, ECCV 2016). Convolutional colour prediction.
- Anime Converter — AnimeGAN v2. Generative adversarial style transfer.
- Face Restore — GFPGAN (Tencent ARC). Generative facial prior.
- Photo Healer — Real-ESRGAN super-resolution + morphological inpainting + GFPGAN.
- Portrait Mode — U-2-Net for segmentation followed by Gaussian blur composition.
- Object Detector — Faster R-CNN with MobileNet V3 backbone, trained on COCO.
Every tool's page links to the relevant guide and includes citations to the original research where appropriate.
Privacy, briefly
Your image is uploaded over HTTPS, processed in a temporary directory on the inference server, and the result is streamed back to your browser. The temporary copy is deleted immediately after the response is sent. There is no logging of image bytes, no retention, no sampling for training, no sharing with third parties. The full privacy policy goes through the details, including what Google AdSense does and does not do with cookie data.
Contact
You can reach me at contact@image-studio.net. I read every email but do not always reply quickly — replies typically come within a few days. Feedback, bug reports, suggestions for new tools or guides, corrections to the guides, and general comments are all welcome.
I cannot provide image-editing services, take requests to process specific images, or offer technical support for problems with your input image. If a tool produces a result you don't like, the tool's page has detailed notes on what each model handles well and what it doesn't.
A note on AdSense
The ads are how the site stays free. They are restricted to specific placements between content blocks and are never overlaid on the tools themselves or on the guide articles. I don't track which articles convert, which placements click, or what individual visitors are doing. AdSense itself uses cookies to personalise ads — Google's documentation explains this, and the privacy policy links to the relevant settings if you want to opt out.
If you find the site useful, the most helpful thing you can do is leave the ads alone (whitelist the domain in your ad blocker, etc.). It is a small ask in exchange for unlimited use of the tools.