Will Stable Diffusion 3 accurately generate ‘blue grass, green sky’ in 9 out of 10 images before July 2024?

Technology AI

Basic

Ṁ2.0k

Jun 30

16%

chance

ALL

This market will resolve as 'Yes' if, before 2025, either I or, in the case of my inactivity, a trustworthiness individual, confirms that Stable Diffusion 3 accurately generates 'blue grass, green sky' images in 9 out of 10 attempts.

Fine-Print Details

The evaluation standard is based on a clear, casual observation of predominantly blue grass and green sky, with minor color imperfections allowed (less than 5%).
Please check my previous markets before betting on this market.

Get Ṁ600 play money

Related questions

Will the Unstable Diffusion team actually deliver their product as described before the end of 2023?

30% chance

Will Stability AI fold by 2028?

84% chance

Will someone figure out how to train decent Stable diffusion models for <$10K USD? (2024 end)

Sort by:

@Soli please resolve

will do this weekend. thank you for the ping

Just got this image out of it for the prompt: "An alien world with blue grass in a field. The sky is a deep green, with blots of greener clouds."

Variations of the prompt also work.

@Soli in the YES market you didn't pass the prompt exactly. What sort of prompt engineering is allowed?

"blue grass, green sky" is continuing to give very mixed results, like this:

i think simple prompt engineering should be allowed since we allowed chatgpt to change the prompt slightly when testing dall-e but nothing super hacky

edit: this is being discussed right now

the first two images you shared qualify, the last two not sure

Hmm, I assumed it wouldn't be allowed since the original DALL-E 3 market said "The prompt should be copied and pasted as "blue grass, green sky," with no extra specifications allowed."

I guess it depends on what you're trying to test. If it's about usefulness, then an internal "prompt optimizer" is just a useful feature, while having to carefully fiddle with the prompt is a big waste of time and compute. If it's about the raw intelligence of the image model itself, then obviously you need to use similar prompts (either similarly optimised or unoptimised) going to the image model.

i am open to any definition as long as we stay consistent with everything i have communicated before and with previous markets (if possible), for now no one should be placing new bets

you are actually right, i checked the first market i created and the prompt was sent unmodified to dall-e so same applies here

DanboughtṀ30NO

In my testing it's not even hitting half success rate.

They could update it before the end of July, the results I'm getting from the API are pretty weak so far though.

bought Ṁ50 YES

Betting on almost no knowledge save for the cover image here lol

@jBosc I just found out this wasn't automatically generated (i.e. by whatever model Manifold uses), but regardless, if DALLE-3 passed the test in December surely Stable Diffusion 3 would pass it now.

@jBosc it was generated by Dall-E 3 which

i believe is the same model Manifold uses