🎨 OpenAI has released ChatGPT Images 2.0 - a new generation of image generation.
The main update is accuracy and control. The model has significantly improved its ability to handle complex scenes with many details, readable typography, and realistic composition. Examples show posters, magazine spreads, manga pages, handwritten notes, and infographics - all with properly laid out text, which used to be a weak point for any image models.
There’s a serious leap in photorealism: cinematic portraits, night shots on film with flash, documentary street photography - all at a level where it’s already hard to distinguish from a real camera. Plus, it works with a stylistic range: photo, illustration, manga, pixel art, comics with consistent characters across several panels.
The most interesting architectural feature is the thinking mode for images. The model can now research, reason, and pull relevant information from the web before generation. So, for a request like "create an infographic based on the latest article X" or "draw the current OpenAI merch," it will first find the data and then layout the visual. Image generation is no longer a separate tool but becomes an extension of the reasoning model.
They’ve also improved production readiness: flexible aspect ratios (from banners to vertical for mobile), character reference sheets with consistent poses and emotions, print-ready layouts with bleed/trim/safe margins. This is no longer just a "pretty picture," but a ready asset for a designer or marketer.
They’ve significantly improved non-Latin scripts: Japanese, Arabic, Korean, Devanagari, Cyrillic, Bengali, Chinese. You can create ready advertising layouts in Korean or manga in Japanese without artifacts in the characters.