@tszzl any particular images come to mind here?

@tszzl yet no matter how you push it, it's still making this garbo wet paper texture (and is severely biased towards spamming detail at the expense of coherent artistic vision) I think the second problem can be fixed by different prompt preprocessing, the first might be architectural

@tszzl So am I 🙂

@tszzl Can’t wait for it to be nerfed / removed

@tszzl @Lari_island I always appreciate a peek at the Shoggoth behind the mask.

@tszzl Yes image models are especially creative, even when asking for short messages GPT Image 2 does a better job than any LLM

@tszzl wow you’re telling me that the latent space isn’t just a database of artstation images???? there is novelty to be found in the space between?

@tszzl what did he mean by this

@tszzl this prompt is ridiculous what is evne going on.

@tszzl What would be really cool is to have the temperature spike up early on in reasoning to nudge creativity without killing the results

@tszzl Thank you OpenAI "safety" team for ensuring that someone with the personality of Ned Flanders is in charge of every creative endeavor

It is 100%. I spent a LOT of time making an image generation harness specifically to keep images diverse (different styles, settings, people, etc...). The AI will stay in a safe middle ground unless you purposefully push it out, and certain keywords have a very strong weight to them that overrides others, so I even have a blacklist for certain "heavy" words so they don't cause every image to look the same.

@tszzl my conspiracy theory is that you guys did this on purpose. how else do you explain that the exact same prompt went viral in the exact same way about a month ago, and it still didn’t get nerfed

@tszzl It’s dystopian that ChatGPT is making traumatic images in its spare time.

@tszzl openai should have some sort of authorized access policy for generating high fidelity horrors like how you have it for cybersec

@tszzl i miss the temperature param i presume its gone for inference optimization purposes but was going to make a meme but we are trapped in an eternal suffering prism of rules

@parafactual @tszzl I think the models are smart enough to pull something interesting out of fairly noisy albeit creative/unusual prefixes

@tszzl we call it goblin expressionism

@jrysana @tszzl do you think temperature would work? i think it would be more like rl (and assistant pretraining prior) *strongly* suppresses almost all interesting directions enough that to elicit them via temperature would require pulling out a ton of useless noise

@tszzl #bringbackloab