ChatGPT Images on X

ChatGPT Images 2.5 is the name people use. For 37 of 55 people here it is simply what they make images with, and what they describe is a step in a bigger job rather than a standalone toy: character sheets — front, side, back, close-up — fed into video tools, four-view assets for game models, a landing page design handed on to Codex, explainers built with images plus layout and animation. Several post the prompt next to the result.

13 people are unqualified fans. "Quality of ChatGPT Images is just nuts", one writes; another pushes graphic designers toward it. Beyond that the section thins fast. 2 people use it for drafts and briefs and keep human review in the loop, with a design guide warning never to publish AI files as official brand assets without checking source files.

3 people have written it off, one of them calling it "completely useless for work" over what they see as constant censorship.

The look is the most specific complaint: output with "that unmistakable AI aesthetic", excessive detail, bokeh and polish, spottable "even on random restaurant menu boards", or a filter people call a "piss filter" and, elsewhere, a melted-wax finish. One says the look of the watermark, once noticed, cannot be unseen.

Characters and likenesses are the other top gripe — faces, lineups and bodies that drift from what was asked, including a prompt that asked for eight characters where seven hit a limit without more specific instructions, and a user annoyed that a chest was enlarged against instructions.

The rest is thin but pointed. Prompts will not hold a logo steady, so people say reference or supplied logo files are required. Two users hit guardrails, one unable to generate "a picture of an angry person". Two mention watermarks or AI metadata, one stripping the metadata with another tool. Two report output going off — "drunk today", timeouts, muddy images. Two hit limits or tools that won't call it, one turning to Google. And two flag false images of political candidates, one citing 16 of 20 attempts.

Nano Banana (and Nano Banana Pro) is the name that recurs, and the verdicts split both ways. One calls ChatGPT Images 2.5 "definitely the better image model in the game right now" while still keeping a favourite Nano Banana template; another rates it "500X better". Against that, one says an older Nano Banana Pro beats it in photorealism even with reference images, nailing the "iphone look" where ChatGPT Images keeps its AI sheen; another says angle instructions land more easily in ChatGPT Images 2.5. Others pair the two, using either for character sheets and skin-texture work.

2 people bring in Codex's built-in image generation: used beside ChatGPT Images for covers and edits where the text was checked character by character, and one finding more noise and grain there and asking whether it is still the older model.

Midjourney appears once, called inferior and "basically free" in ChatGPT's favour. Google once, used after a user ran out of images here. Canva once, in a same-prompt duel. One turns to other tools because their characters hold across generations.

The section holds 1 post from 1 person, so there is little to generalise from. The suggestion is about model routing: assigning a faster engine or a more detail-focused one to each task, rather than "paying the computational cost of the heavier model on every generation". The person points at what they take as precedent — within ChatGPT Images, a speed-first variant and a detail-seeking one — and argues the same structure would let the maker match the engine to the job.

Nothing else is asked for here: no request for cheaper pricing, better text, fewer guardrails or a particular feature appears in this section, and the posts filed under problems are complaints, not asks.