Hacker News Reader

Comment by jamilton | Hacker News Reader

It's interesting that the outputs still vaguely resemble the subjects of the prompts. It implies to me that there's still a little bit of the characters in the dataset, but not very much.

If you look up "Pikachu" in Adobe Stock [1] there actually are a lot of results. They're mostly figurines, pokemon cards, quokka, and Machu Pichu. Very few direct illustrations. And of course there might be filtering used on the dataset before it was used to train the model, a lot of the images that have Pokemon imagery are marked "Editorial use only".

1: https://stock.adobe.com/search?filters%5Bcontent_type%3Aphot...

View on HN

My uninformed speculation is that this might be an artifact of the embedding layer (e.g. CLIP), not the image training data.

Presumably, the dataset for training the embedding layer (which is trained separately and then fixed) is not stripped of copyrighted content. So it will have learned that "Pikachu" is related to words such as "yellow" and "rat".

Therefore, even if the image dataset didn't have a single picture of Pikachu, the image generator will still likely produce a yellow rat. Just one that doesn't actually resemble Pikachu.

That seems plausible to me. I assume if Adobe felt like even these results were too legally risky they could create a new embedding layer with fewer copyrighted concepts (so Deadpool would always output a spooky looking pool). I don't think that's likely though.

> It implies to me that there's still a little bit of the characters in the dataset, but not very much.

I suspect there are a lot of the characters in the training data, but (postulating) there might be a post-generation adversarial selection step where anything that is recognizable as a someone else's IP gets filtered out of the results, so anything sufficiently Pikachu-like won't be shown, leaving behind the vague yellow monsters.

I would think attempting to filter it out after generation would be harder and less effective.

ML image recognition predates diffusion models and is more mature - I'd be shocked if it's not employed in some form.

Such a system would be more robust against incidental infringement of IP not mentioned in the prompt. e.g. A prompt for "Theme park" won't create Mickey Mouse in the background

Reply on news.ycombinator.com