Developer known as jjcm drove away ran the new OpenAI ChatGPT Images 2.5 model through a set of interface generation tasks and compared the result with the previous version of gpt-image-2.
The general conclusion is that the new model works noticeably better with the reference images provided to it. Some complaints about the previous version have been resolved, but new ones have appeared.
The first test was the interface of the agent-based development environment in the style of Warcraft 3. The difference in working with samples was most easily revealed here, since gpt-image-2 could not accurately match the desired shade of the blue button, and version 2.5 reproduced it without edits.
The community received the result with mixed feelings. Some readers noted that the picture was close to how they remember the game, others criticized the details.
One of the commentators:
I can forgive the chaotic placement of chains, but I can’t forgive the soulless character icon from the WoW era. The WC3 interface would have used a hero portrait.
The second test was a page of a fictional politician with the face of Glenn Powell. Here, the model’s ability to change the character’s pose while maintaining a recognizable appearance was tested, and 2.5 coped well with this task. In the previous version, on similar layouts, the skin acquired a characteristic overcooked look with excessive contrast and sharpness.
The third test was a website for a ramen establishment with cyberpunk aesthetics. Dark interfaces used to most strongly provoke the same “overcooked” image, and on this front the new model looks cleaner.
It was precisely on this layout that the main regression was revealed. The model renders small decorative symbols worse, as can be seen from the turquoise lines on the bottom right of the bowl of ramen, which turned out blurry and uneven.
The side effect of generating interfaces was noted separately. The page of the fictional politician in the discussion was called impeccably smooth, but completely bland and fake, which turned out to be rather on target for the layout of a political website.
All three tests were carried out by uploading our own reference images, and not using a text description.
