Introducing GPT Images 2.5: Faster, More Realistic AI Image Generation
GPT Images 2.5
A major update has arrived: quality and realism have been improved, it now produces more natural lighting and rich textures, better preserves the features of people and objects from reference photos, and subject recognizability has become noticeably higher.
They have significantly “improved editing”, claiming that the model changes only what it is asked to, without touching the rest even with complex objects and backgrounds, and supposedly there is no more degradation between shots. I put that in quotes because recently I couldn’t simply ask GPT Images to change the lower part of the clothing in a photo — it was just wrecking the face completely))
Generation latency has been reduced by up to 50% compared to Images 2.0 + concepts can be iterated almost twice as fast.
It handles complex visual requests better, more accurately reflects real information, supports complex layouts including transparent backgrounds, and more accurately conveys artistic style.
New tools
Sketch — you can draw in the chat (just type @Sketch): then make some sketch and generate an image from it.
Templates — a quick start for popular formats: posters, merch, flyers, product photos.
Comments on images — you can point out precisely what to tweak.
API news
Two new models have appeared in the API:
GPTImage2.5 Flare — higher quality than GPTImage2 with 50% lower latency, intended for mass generation.
GPTImage2.5 Sunburst — premium level with maximum control during editing (better for already finished projects).
Images 2.5 is already being rolled out to all ChatGPT, ChatGPT Work, and Codex users on all plans — desktop, mobile apps, and web. Both API models are also already available.
Regarding safety, OpenAI retains prompt and image checks, C2PA metadata, and invisible watermarks to identify generated content.