ChatGPT Images - v2.0

Update 9/9: ChatGPT Images 2.5 is now available for on-site generation, in both the Flare and Sunburst builds.

ChatGPT Images 2.5 is OpenAI's latest text-to-image and image-editing model, released September 8, 2026 as the successor to Images 2.0. It brings sharper detail, stronger style adherence and more control over edits, and it is noticeably better at working from reference photos - subjects you supply are more likely to stay recognizable when you move them into a new setting, style or composition.

Originally posted at https://openai.com/index/introducing-chatgpt-images-2-5/. All credit for the model goes to OpenAI. Civitai is hosting it so creators can run it on-site without an API key.

Two builds: Flare and Sunburst

OpenAI ships 2.5 as two API models, and both are available here as separate versions on this page.

• v2.5 Flare is the default choice for most work. It produces higher-quality images than Images 2.0 at roughly half the latency, which makes it the one to reach for on creator and social content, product shots, and anything you plan to iterate on.

• v2.5 Sunburst is built for premium visual workflows that benefit from tighter control across edits - production-ready campaign creative, polished product imagery, and anything where you are refining the same image over several passes.

They take the same controls, so switching between them is just a version change. Start on Flare and move to Sunburst when an image needs the extra control.

What is new in 2.5

OpenAI describes 2.5 as producing "more natural lighting and richer textures" and being "better at preserving the subjects in your reference photos." In practice the gains show up in three places: fine detail holds together instead of dissolving, a requested style is followed rather than approximated, and edits are more reliable across several turns so a series of adjustments does not drift away from the original image.

Speed is the other headline. Flare generates higher-quality output than Images 2.0 while cutting latency by up to 50%, so the model is now practical for high-volume and rapid-prototyping work that 2.0 was too slow for.

Generating on Civitai

Both 2.5 builds support text-to-image and image editing on-site. Attach one or more reference images and the generator switches to edit mode, using what you supply as the basis for the result. Quality is selectable at low, medium and high, and the available aspect ratios are 1:1 (1024x1024), 3:2 (1536x1024) and 2:3 (1024x1536).

Like the rest of the OpenAI family on Civitai, 2.5 is a hosted API model: there are no weights to download, and it does not take LoRAs, negative prompts, samplers, steps or CFG. Write what you want in the prompt and the model handles the rest.

About ChatGPT Images 2.0

The previous generation is still on this page as its own version. Originally posted at https://openai.com/index/introducing-chatgpt-images-2-0/.

Images 2.0 was a step change in detailed instruction following, in placing and relating objects accurately, and in rendering dense text. Its sense of composition means results feel less AI-generated and more intentionally designed, and it is accurate across languages, using expanded visual and world knowledge to fill in the gaps so you get smarter images with less prompting.

Greater precision and control. Images 2.0 can render the fine-grained elements that often break image models: small text, iconography, UI elements, dense compositions and subtle stylistic constraints.

Stronger across languages. It brought significant gains in non-Latin text rendering, particularly Japanese, Korean, Chinese, Hindi and Bengali, producing images where the text is not only correct but reads coherently as part of the design.

Stylistic sophistication and realism. It captures the defining characteristics of photographs, including the small flaws that add realism, as well as cinematic stills, pixel art, manga and other distinctive visual languages.

Real-world intelligence. Images 2.0 has a knowledge cutoff of December 2025, which matters for explainers, educational graphics and visual summaries where correctness counts as much as aesthetics.