Permission request: using this model on a free Norwegian site
Hi Smogy, @Smoggy
I'd like to ask permission to use your model, and I want to be completely honest about who I am, what I've built, and how I'd use it.
Who I am: Not a developer. I'm a hobbyist in Norway who has built a handful of websites. This project started after a conversation with my daughter about how you can tell a real photo from an AI-generated one. I didn't have a good answer for her, and that bothered me enough that I started building something. I've built it with a lot of help from Claude (an AI assistant) — I genuinely couldn't write most of this code myself, and I'd rather say that plainly than pretend otherwise.
What it is: bildesjekk.no — a free Norwegian service where you upload an image and get an assessment of whether it looks real, edited, or AI-generated. It checks metadata, C2PA Content Credentials, and runs a visual analysis. What makes it different from most detectors is that it explains its findings in plain Norwegian instead of just printing a number, and it's built around a firm principle: never tell someone their real photo is fake.
What I tested: I ran your model (the ONNX conversion published by onnx-community) on 17 images — 11 real photos straight off my own phone, and 6 AI-generated images.
Real photos (11): every single one scored 0.0%. Not one borderline case. That included night shots, backlit sunsets, blurry dog photos and heavily compressed Snapchat images. Zero false positives.
AI images (6): caught 4. It correctly flagged a polished AI "office portrait" at 93% — an image that had completely fooled the vision model I currently use, which rated it as probably a real edited photograph. That's exactly the failure case I've been unable to solve, and your model solved it. It missed a recent ChatGPT-generated landscape, which I assume is simply outside what it was trained on.
That combination — never accusing a real photo, while catching the cases that fool language models — is precisely what I need. It would be an additional signal in my results table, clearly labelled as probabilistic, never presented as proof.
My question: The license is CC-BY-NC-4.0. My site is free and has no ads, but it does have a "Buy Me a Coffee" donation link to cover hosting and API costs, and I've considered adding ads or a paid tier at some point. I honestly don't know whether that crosses the line into commercial use, and I'd rather ask you directly than guess and get it wrong.
Would you be willing to let me use the model on bildesjekk.no? I'd credit you and the model prominently, link back to your Hugging Face page, and be explicit on the site about what the model can and cannot do.
If the answer is no, that's completely fine — I'd rather know now than build it and find out later.
Either way, thank you for publishing this. Of everything I've tested, it's the only thing that got that portrait right.
morkerom
bildesjekk.no