@Hazel what model is that and how did you prompt it?

@ProjectVictory lumalabs, used an iterative version of their new model (re-prompted dozens of times until the output was perfect)

@ProjectVictory it would be trivial to create an API that did this automatically, in essence, making a much improved model.

Still, this was cherry picked. The king/queen is still the hardest part.

@Hazel Did you use a fixed series of prompts? If not, how would you make an API that does this automatically?

@MaxMorehead yes, same prompt over and over. If I wanted to make some mana, I could easily build this before the end of the year. It’s easy to repro.

@Hazel oh, would have taken a couple more iterations, to get the queen in the right place. I’m a callable human in the loop lol.

To be clear, you have to verify it's correct?

@Shump How would this resolve if it's possible to build a scaffolded system that generates a chessboard (e.g. calling DALL-E multiple times, using GPT-4o to verify whether the image is correct). Would it change if there's more purpose built parts to the scaffolding (e.g. taking subsets portions of the image and using specific prompts to verify those)?

recraft v3 does not seem to be better than flux

@TobiasWegener
I think we are getting pretty close with Flux

Problems:
there seems to be a rug, and both sides are white.

The figures seem quite good now.

@TobiasWegener Also the closer queen is a bishop and a bishop is a smaller bishop

@ProjectVictory yeah you are right and the strange line in front of the queen, a lot of small mistakes. Intersting how hard it is to see many of them.

I think there's a difference here between generating the starting position as an image and generating a random mid-play board but ensuring that it could have come from a real start position. The latter is a much more difficult task

I agree, but the title here is slightly misleading, the description clearly calls out they want a starting position

I've been trying to cue the model into producing a diagram, since that's presumably easier, but it's not quite getting there. I think the problem is very similar to producing text, if you think of chess pieces as symbols and chess boards as phrases.

https://x.com/LukeASalamone/status/1820299359185211901?t=moIpHPN-SCu7Zw_WSom3CQ&s=19

The Flux model is much closer than DALL-E, but not quite there yet (maybe with the right prompt though, I just used "chess board starting position, view from above")

"chess board starting position"

this one's quite good, though of course the knights are very broken

It's quite wrong. The board is rotated incorrectly. The bottom left corner should be white

@AndreiVlasenko close but white queen not on her color

Does GPT-4o count? If so, this is going resolve to yes.

@Cosmic1 do we know if gpt4 is using a new image generator? afaict it's the same interface to dalle3 as before. There is no new image endpoint available via openAI api.

@Ernie GPT-4o is an image model, but it does not currently output images. It will in the future

@Cosmic1 Wrong

@diadematus It’s literally not wrong. “We’re announcing GPT-4o, our new flagship model that can reason across audio, vision, and text in real time.”

https://openai.com/index/hello-gpt-4o/

“…generates any combination of text, audio, and image outputs.”

@Hazel how does GPT-4o being able to "reason across ...vision" resolve "generate correct images of a chess game" as YES?

@Cosmic1 In ChatGpt you have to be careful that GPT-4o doesn't use the python code interpreter. With that it can easily generate a perfect image but it is not what the question asks for.

@Hazel Right, but once GPT 4o image gen comes out… it’s gonna be quickly solved

@Cosmic1 Yes GPT 4o count, with or without DALL-E. As long as it is a general purpose generative model that makes images from text only, it counts. Images generated from code don't count as they're not generated by the AI