ChatGPT Images 2.0 adds dense text and multi-image generation

• OpenAI says the launch post adds denser text handling, stronger instruction following, a reasoning-backed thinking mode, and multiple images from one prompt, while the API announcement frames the same model as production-ready for layouts, edits, and on-brand assets.
• According to the HN discussion roundup, early developer attention went straight to the gpt-image-2 model card and pricing, which is a clue that this shipped as both a ChatGPT feature and a developer surface.
• In the main HN thread, commenters kept stress-testing the same few use cases: comics, dense text, layout-heavy prompts, and whether the model is now good enough for product-photo work.
• One of the clearest early workflow signals came from vunderba's HN comment, which said gpt-image-2 held continuity across custom multi-panel comics, a task that used to break older image models fast.
• Attribution anxiety showed up immediately in the HN thread, where commenters raised C2PA, while OpenAI's system card says Images 2.0 ships with C2PA metadata plus an imperceptible watermark for provenance checks.

You can read OpenAI's launch post, skim the API model page, and go deeper with the official prompting guide, which gets unusually concrete about size limits and text-heavy layouts. The community immediately latched onto the HN thread, where comics continuity and source attribution came up almost as fast as the headline features. OpenAI's own system card adds one more useful detail: the realism jump is strong enough that provenance tooling now sits in the center of the launch, not the footnotes.

The API announcement says the model is aimed at assets that need to be readable, localized, on-brand, and usable without heavy cleanup. It specifically calls out diagrams, infographics, charts, posters, and comics.

The official prompting guide fills in the practical part:

• gpt-image-2 is OpenAI's recommended default for text-heavy images, photorealism, compositing, and identity-sensitive edits.
• Resolution is flexible, but both edges must be multiples of 16.
• The max edge is under 3840 px.
• Aspect ratios can stretch to 3:1.
• Anything above 2560 x 1440 is treated as more experimental.

That mix is why the release reads less like an art toy and more like a layout engine with an image model attached.

According to the launch post, thinking mode can use reasoning, web search, and tool use before rendering. The API announcement adds a more grounded description: reasoning models can research inputs, transform them, generate variations, and self-check outputs.

OpenAI also says Images 2.0 can generate multiple images from a single prompt. That matters because coherent sets, not just one-offs, are the real bottleneck in decks, campaign assets, storyboards, and product flows.

The developer rollout also landed on day one. The API model page lists gpt-image-2 as the main snapshot, and the community announcement says it was available immediately in the API and Codex.

The most creator-specific test in the evidence pool is comics. In the HN discussion digest, commenter vunderba said gpt-image-2 handled custom multi-panel comics well and stayed faithful to continuity across panels.

That lines up with OpenAI's own positioning. The API announcement includes comics in its list of stronger structured outputs, alongside infographics and posters.

For artists and marketers, continuity is the threshold feature. A model that can keep characters, objects, and text stable across panels stops behaving like a one-shot generator and starts behaving like a rough layout partner.

A separate thread inside the HN discussion focused on whether the realism jump pushes image models into product-photography territory, and whether that makes unlabeled assets harder to trust. One commenter explicitly raised C2PA as the standard to watch.

OpenAI's answer is in the Images 2.0 system card. It says the realism increase could enable more convincing deepfakes without safeguards, so the release includes prompt classifiers, image classifiers, post-generation policy checks, C2PA metadata, and an imperceptible content-specific watermark.

That provenance stack is new information worth paying attention to because it tells you what OpenAI thinks changed most: not just prettier images, but images realistic enough that origin has to ship with the product.