Ran with latest ChatGPT Pro with Extended mode. I believe the final model (ChatGPT Images 2.0) is distinct. It seems clear to me you get better images with Pro than Thinking, but the details of how they work together isn’t fully explained anywhere AFAIK.
@goodside
-
Reports of prompt engineering’s death mildly exaggerated
By
–
Reports of prompt engineering’s death have been mildly exaggerated:
-
Code-generated reference image yields any shape like ceramic nonagon coaster
By
–
Telling it to use a code-generated reference image not only gets you to a regular nonagon, it gets anything you want shaped as one—e.g. this ceramic drink coaster, made in Thinking Extended mode:
-
ChatGPT Images 2.0 generates cake photo with SVG that renders another cake
By
–
ChatGPT Images 2.0 (Pro) generates a photo of a cake decorated with SVG that when transcribed to a file renders another cake
-
ChatGPT Image 2.0: Man returns canvas print of himself being photographed at HomeGoods.
By
–
ChatGPT Images 2.0 (Pro) generates an amateur photo of a man returning a decorative canvas print to HomeGoods because the print is actually just the same candid photo of him being taken at that moment and that doesn’t make any sense.
-
ChatGPT Images 2.0 generates a cake f(x)=x^3-3x^2+2x
By
–
ChatGPT Images 2.0 (Pro) generates an f(x)=x^3-3x^2+2x cake:
-
Knowledge of Wikipedia URLs for numbers without searching
By
–
It surely knows the Wikipedia URLs for numbers without searching. I’m not sure it searched anything.
-
Images 2.0 Pro one-try success vs ten Thinking failures
By
–
I know but that was before the Images 2.0 release. The image at the top of this thread was made using Pro in one try, and before I had Pro I did ~10 tries with Thinking and they all had unacceptable errors. I’m pretty sure it’s different.
-
Extended thinking produces helper images for multimodal input; Pro beats Thinking Heavy.
By
–
Yes. It’s counterintuitive but extended thinking actually helps now because it can code-gen “helper images” in the CoT to use as multimodal input for the final generation. Pro is clearly a better image generator than even Thinking Heavy though.
-
ChatGPT Images 2.0 generates sink with Wolfram rule 182 ceramic backsplash
By
–
ChatGPT Images 2.0 (Pro) generates a sink with a the first 16 states of Wolfram rule 182 as a ceramic tile backsplash. I.e., Sink-182.