Hacker News Reader

Comment by executesorder66 | Hacker News Reader

AIs are still not able to understand negations.

Try "ramen without egg" or "ramen with no egg" and it will show ramen WITH egg.

Or "man without striped shirt" will give "man WITH striped shirt"

It's not trained for it, because that use case is handled differently. It would be mostly a waste of time to train the concept compared to other things you want to achieve. Instead you put things you don't want in the negative prompt. This example doesn't expose the option, but you can try it here for a different model: https://huggingface.co/spaces/gokaygokay/Kolors

Set the seed to 0 and prompt to "man in a loud shirt" - you get flowers. Sweet the negative prompt to "floral shirt" - no not flowers.

Sentence processors can definitely understand negation, (any non-trivial LLM can) but it would be a waste of time to train that in the image generators -vs- making other ideas better.

Why can't an imagen generate run a tiny little automatic text-to-text rewrite first, to apply these special linguistic rules?

some do. but generally people just learn to use the tools.

Apps / interfaces to those models can totally add that. But it's not necessary to add that to the model itself.

That’s what negative prompt is for. Stable diffusion also isn’t like llms. LLMs certainly understand negation.

I did mean AI's in general, so I have edited my original post.

> That’s what negative prompt is for.

This is what I mean by it "not understanding negations" You need whole separate prompt, just to say you want e.g. "ramen without egg" instead of just saying it in a single prompt that it understands.

You are not correct. LLMs understand "ramen without egg". Image gen models generally do not as this is not how images are described.

If you want to generate ramen without egg, you'll want _negative weighted_ prompts. "eating ramen, (egg:-1)"

Negative weighted tokens don't do what you think they do. Sometimes they act like a negative prompt, other times they don't.

Likewise, zero-weight tokens don't act like the token is absent from the prompt.

They generally do. It is difficult to differentiate "negative tokens don't do that" and "prompt adherence is shaky in general".

It's fully possible that the image model draws eggs in ramen but it doesn't know that the egg is an egg and therefore any attempts to interact with it via the egg token are futile. Generally speaking though thing:-1 should reduce the presence of thing for well understood concepts. It's a better tool on second pass alterations of an image.

You're not correct about AIs in general. Both chat LLM models and sentence embeddings can handle negation just fine. (Ask any chat "what clothes would a person wear if they weren't wearing a hat") Here it was simply not trained for that purpose. Maybe it wasn't worth it, maybe the creators thought that the negative prompt is enough, maybe the time was better spent on other examples. They way, it's not AIs in general, and it's not a tech limitation.

>AIs are still not able to understand negations.

AIs are able to understand negations, just ask an LLM a question. Text-to-image models are the ones that struggle the most with this, they usually do not have a very nuanced understanding of text.

Reply on news.ycombinator.com