Tested image generation with GPT-5.6 Sol (in ChatGPT).

Now we’re doing the real hype test I’m using the image-generation skill and choosing the concept myself—beautiful, cinematic, and distinctly strange enough that it shouldn’t feel like generic fantasy wallpaper.

ChatGPT-5.6 Sol, just two words prompt…my two “fave” words…hard to guess..not.

Tested on gpt-image-2, which is what you are getting anyway, despite a new middleman to pull a tool lever having been attested here recently.

Now we’re doing the real hype test :smirking_face: I’m using the image-generation skill and choosing the concept myself—beautiful, cinematic, and distinctly strange enough that it shouldn’t feel like generic fantasy wallpaper.
- Keep any in-image text incidental and natural, with no obvious labels, carriers, nor explanatory wording.

2240x1600/low

2240x1600/medium/tweaked input

_j:

Tested on gpt-image-2,

Haha I still now it is gpt-image-2, but it uses image skill, so the output is more different than directly with gpt-image-2 or using gpt-5.5 as “middle man”, but I like your results

Very good analysis @_j
Giving us base level examples to use is perfect.
You guys should do a separate chat each model same prompt starting with dalle

mitchell_d00:

starting with dalle

That could only serve as a reminder of what has been lost.

Mating Dance

An abstract pencil and watercolor drawing oof two red-tailed hawks doing a mating dance.

From my side porch, just saw two hawks doin the dance…

A photorealistic image of an astronaut riding a horse

DALL-E 2 (By OpenAI, 2022-11)

gpt-image-2 (2026-07)

(the second row gives quality:low, in comparison to quality:medium)

DALL-E 2 didn’t write its opinions into your pictures, though, which you should now enjoy.

I have not played with art in a while don’t “send exact” still work? If not that’s a shame. Is @Daller still doing art? @polepole?

There is really no concept of a middleman that you have to talk your way into sending the clear text you want for CLIP embeddings now - gpt-image itself is large language model technology, knowledgeable, and which can make inference while the model itself generates via multimodal understanding.

This might not go over well on DALL-E 3

# task
Via generalization of Issai Schur's inequality and functional extension to more variables, algebraic spaces, and/or Hermitian matrices, demonstrate a new application that is novel to mathematical sciences.

# response
linear image output

5488 “tokens” of imagery output…