Bottom line is that It appears that gpt-image-2 is not entirely trained for sprite sheets - the subject of this thread.
Rather the ones who find AI model that can tell accurately “is it left leg of the person in the bottom right of the image” can become a millionaire in a month…
Maybe this guy can fix it:
Great discussion, I’m also working on this problem and landed on a similar “solution” to OP. Only issue is its tough to find pose references for northward and southward movement. Anyone have any luck with this?
Hi and welcome to the community!
I previously used a video generator to create a full animation. This method can also be used to create and extract reference images.
Codex now has pets, that uses gpt-image-2 to create its sprites. Perhaps this skill might hold some valuable tips.
aka, “make free to bill with any API key environment variable found”, and “run and persist arbitrary AI-created code”. 10+ images with image input billed in agentic experimentation, anticipating failure in image alignment inspected with vision and failure in faked transparency with image retry scripts also.
All for bespoke code that already exists.
Point your cheapest image setting at a 1k bitmap action sprite pose such as awake.xbm from the source code with a new reference character. Cheap is not gpt-image-2, because you automatically pay 4k or 6k additional tokens for each input image at increased prices even at the smallest resolutions sent besides the vision tokens.