The Pulse
OpenAI Ships ChatGPT Images 2.5 With Faster, More Precise Edits
OpenAI has released ChatGPT Images 2.5 with improved reference-photo fidelity, multi-turn editing, sketch-based generation and up to 50% lower latency. The update also adds templates, image comments, shared prompts and two new API models.
AI.info Team ·
When Axios tested OpenAI’s new image engine with a photograph of a black cat named Raven, the model turned the animal into a film-noir detective, a stained-glass subject and an anatomical diagram while preserving what the tester called Raven’s “wonderfully unimpressed expression.” A separate tattoo-editing test let the user circle the area to change, request lighter elements and refine the design over several turns.
Those demonstrations capture the focus of ChatGPT Images 2.5, which OpenAI released on September 8. The update targets a persistent weakness in image generators: making substantial changes without losing the identity, composition or details of the original image.
OpenAI says people create more than 3 billion images each week across ChatGPT Images and its GPT-Image models in the API. Images 2.5 is now rolling out to ChatGPT, ChatGPT Work and Codex users on the web, desktop and mobile, across all tiers.
Axios’s hands-on testing found that the new editor handled reference images and targeted revisions more effectively than earlier versions, including edits made by marking up part of an image rather than describing every change in text.
Reference photos get more consistent treatment
OpenAI says Images 2.5 produces more natural lighting, richer textures and more recognizable subjects when it works from reference photos. Users can place a person, pet or other familiar subject into a different setting, visual style or composition while retaining more distinctive features.
The model also aims to preserve the rest of an image when a user changes one element. An editor can update a product, background or piece of copy while keeping the subject, composition and surrounding visual treatment intact. That matters for repeated creative work, where rebuilding an entire asset after every revision can introduce unwanted changes.
“What impressed us most about GPT‑Image‑2.5 Flare is how well it understands what not to change,” said Axultan Alimkulov, head of product at Higgsfield AI. “You can make a meaningful edit without losing the character, composition or visual identity of the original image.”
OpenAI also says the model holds onto earlier changes more reliably during longer conversations. Each new instruction is intended to build on the previous version without degrading image quality or undoing details established earlier.
Sketches, templates and comments move editing beyond prompts
ChatGPT Images 2.5 adds several controls aimed at users who do not want to describe every visual adjustment in prose. The new Sketch feature lets users draw directly in ChatGPT and use the drawing as a guide for the finished image. A rough room layout, outfit outline or object can serve as a visual reference before the model generates the completed result.
Templates provide starting points for formats such as posters, flyers, merchandise and product photographs. Users can then add the information, design elements and visual style they want. OpenAI’s help documentation says templates can also accept reference-image uploads when a category supports them.
Users can place comments directly on an image to identify where an edit should occur. They can also share the prompt used to make an image, allowing another person to reuse the idea with different photos and details instead of copying the original asset unchanged.
Flare and Sunburst split the API release
OpenAI is releasing two API models alongside the ChatGPT update. GPT-Image-2.5 Flare is the default option for most applications and is designed for creator tools, social content, product experiences, visual search, prototyping and high-volume generation.
OpenAI says Flare delivers higher-quality results than GPT-Image-2 with 50% lower latency. GPT-Image-2.5 Sunburst is aimed at workflows that need tighter control across edits, including campaign creative and polished product imagery, with longer generation times.
That distinction gives developers a choice between faster, general-purpose generation and a slower model intended for detailed creative work. The models are also available through partner products, including Adobe Firefly, according to OpenAI’s announcement.
Speed improves, but realism raises safety pressure
OpenAI says Images 2.5 reduces generation latency by up to 50% compared with Images 2.0. Faster output is especially relevant to multi-step editing, where users may make several small changes before accepting a final image.
The company’s safety card acknowledges that greater realism can make deepfakes involving real people, places and events more convincing. OpenAI says its safeguards check prompts, input images and generated outputs, using policy checks, image classifiers and a safety reasoning model before an image is shown.
In OpenAI’s adversarial evaluation, unsafe outputs were presented 1.09% of the time for Sunburst and 1.41% for Flare, compared with 1.64% for the ChatGPT Images 2.0 baseline. The company cautions that the test set was designed to provoke policy violations and does not represent ordinary production traffic.
OpenAI says Images 2.5 continues to use C2PA metadata and invisible watermarking to help identify images made with its tools. The release puts the update directly in front of ChatGPT users while giving developers two separate models for production image workflows.