KI-Bildgeneratoren verwenden standardmäßig dieselben 12 Fotostile, Studienergebnisse
https://gizmodo.com/ai-image-generators-default-to-the-same-12-photo-styles-study-finds-2000702012
KI-Bildgeneratoren verwenden standardmäßig dieselben 12 Fotostile, Studienergebnisse
https://gizmodo.com/ai-image-generators-default-to-the-same-12-photo-styles-study-finds-2000702012
12 Kommentare
They needed a study?
Well yeah. Because it can only steal from existing work and it averages everything out to the fewest different categories of a given thing.
My brain can only generate 12 types of nipples and they’re all pink.
Makes sense to me. If you want the most middle-of-the-road, generic looking stuff then yeah AI can create that.
It seems like they tested 2-4 year old AI models, and many of these older models were trained on the same datasets.
> stable-diffusion-xl-base-1.0, segmind-SSD-1B, stable-diffusion-v1.5, and playground-v2-aesthetic
The „base“ version of models are also meant to be finetuned, so they aren’t going to be great for normal use.
> We built iterative feedback loops between Stable Diffusion XL (SDXL; image generation) and Large Language and Vision Assistant (LLaVA; image description), forming autonomous text → image → text → image cycles.
It also looks like they are introducing biases from the VLM models they chose, which is going to have a major impact on the direction their feedback loop takes. And they were only testing both model types with basic settings.
Stable diffusion and Llava are not representative of all AI image generators. Should have tried with DallE and Nano Banana too
And guess what? The more we all use it while paying for it, we’re giving these companies more and more information to use to eliminate the need for us. If there were an argument for a universal monthly dividend or payment from a company this would be it.
I read the original paper and they linked this image of 100 different convergent endings: https://www.cell.com/cms/10.1016/j.patter.2025.101451/asset/932621f9-00d0-4854-aa94-9c9139dcb3df/main.assets/gr4_lrg.jpg
Unclear where the 12 in the Reddit title comes from
There are ways to do advanced mixers where you can produce entirely new styles. This is how some people reproduce the art of dead Manga artists for example.
I Feel Like There Is No Free Image Generator That Is Not Repetetive
B b but I’m an artist! I wrote the prompt to tell it to make an art!!!! /s
Ai is a toy people with no talent love because it makes them feel talented.