News for hackers. Stuff that matters.

A comment on Stable Attribution.

Re:Stable Attribution by GaggiX on Sunday February 05, 2023 @06:49PM (#34671198)
Calling the nearest neighbors of the CLIP embeddings of an image "attribution" feels really misleading, the model has been influenced by the entire dataset it was trained on, just by finding the most semantically similar images does not mean the AI is just using that speficific group of images as references, they probably have almost no influence compared to the entire size of the dataset.

P.S. I'm having fun uploading actual photos and art just to see what the site tell me with confidence "These human-made source images were used by AI to generate this image".

Edit: https://rom1504.github.io/clip-retrieval, this site has always been there to explore the LAION dataset using CLIP image/text embeddings and without the need to mislead the user.

Edit 2: As it's showed in this tweet: https://twitter.com/kurumuz/status/1622379532958285824, they are just using CLIP-L/14 and find the most semantically similar images.

[ Parent | on HN ]
Re:Stable Attribution by __forward__ on Monday February 06, 2023 @06:05AM (#34675849)
Yeah this needs to be higher up, very misleading! To be fair though, they do point it out in their FAQ: "Version 1 of Stable Attribution’s algorithm decodes an image generated by an AI model into the most similar examples from the data that the model was trained with."

To me it actually seems like this language intentionally obscures that they are 'only' using similarity search.

Imo this does a huge disservice to AI communication, providing fake explanations where even cutting edge research has poor insight into these models.

[ Parent | on HN ]
Re:Stable Attribution by ausbah on Monday February 06, 2023 @10:42AM (#34678990)
I don't think this is right. it's possible to fine tune these models with a few pieces from a select artist such that when you say "[art] is the style of [new artist]" you will get new pieces in the style of that artist. those few select pieces have clearly had disproportionate influence on the generated images, even if just via conditioning in the prompt
[ Parent | on HN ]
Re:Stable Attribution by GaggiX on Monday February 06, 2023 @11:59AM (#34680353)
Okay but here we're are talking about the standard SD 1/2 model trained on the LAION dataset, the site only do CLIP retrieval on the LAION dataset, the only thing that makes this website different from Google/Yandex/etc image reverse search.
[ Parent | on HN ]