OpenAI released ChatGPT Images 2.5 on Sept. 8 in the U.S., putting the focus on image editing, reference-image fidelity, and consistency across multiple rounds of revisions.
According to OpenAI, users making repeated changes to the same image should see previously adjusted elements stay in place more often. The company also said edits to a person, product, or background are more likely to affect only the requested area. Generation speed is up to 50% faster than Images 2.0, OpenAI said.
At the same time, ChatGPT gained Sketch, Templates, and image comment features. Users can draw a sketch as a reference or mark parts of an image that need changes. On the API side, developers can use GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
OpenAI said users now create more than 3 billion images a week through ChatGPT Images and GPT-Image models in the API. Images 2.5 is already available to ChatGPT, ChatGPT Work, and Codex users across web, desktop, and mobile.
Editing after the first image is where the update matters most
The hardest part of AI image generation often starts after the first image looks good enough. A user may want to keep the person and only change the clothing, keep the product and only swap the background, or leave a poster intact and revise just one line of text. In earlier workflows, changing one part could alter other parts of the picture as well.
OpenAI placed precise editing near the center of the 2.5 update. The company’s examples include “full body edit” and “multi-city travel ticket,” both shown as step-by-step image sequences to test whether the model can revise a specific target while preserving the subject, composition, and other details.
The travel-ticket example is easy to map to practical use. If only one field needs to change, the model has to identify the exact object to modify while keeping the overall ticket design intact. For product images, ad assets, and branded posters, that can matter more than producing a single attractive image from scratch.
Multi-step editing is another major part of the release. OpenAI showcased “cube rotation,” “travel infographic,” and “birthday candle,” all built around several rounds of edits on the same image. The point of those demos was straightforward: earlier edits should remain in place instead of getting damaged by later changes.
That is also the part of Images 2.5 drawing the most attention.
X user @thesoragirls ran a direct test by drawing a candle in an image, asking the model to keep one flower petal unchanged, and then checking whether edits to other regions would disturb the fixed part. According to her result, the protected petal stayed in place while other parts changed as requested.
Still, repeated editing has not removed all detail-related issues. Japanese animation creator @genel_ai said in practical testing that while OpenAI emphasized image quality across multiple edits, jagged edges and texture changes could still appear as revisions stack up.
That leaves a mixed picture. Version 2.5 appears to improve stability, but image quality loss can still show up in more complex scenes.
Reference images are a second major upgrade
The other big change in Images 2.5 is how it handles reference images.
Users can provide a photo of a person, pet, or another subject, then ask the model to place that subject in a new scene, apply a different visual style, or change the composition. OpenAI said the new version is better at preserving recognizable identity traits while also improving lighting and texture.
The company presented five reference-based cases: “reimagined baby portrait,” “dog in costume,” “photo booth headshots,” “multi-person party composite,” and “making the bed.”
Those examples cover different demands. The baby portrait and photo booth headshots test whether facial traits stay recognizable. The dog example looks at whether the subject remains consistent after a costume change. The multi-person party composite pushes the model to keep several people coherent in one image. “Making the bed” moves closer to everyday editing and asks for concrete changes inside an existing scene.
That kind of consistency matters in workflows built around the same character, product, or brand asset. API users can generate multiple versions from one reference image and cut down on the visible drift that often appears when they start over each time.
OpenAI also used complex scenes to show what the model can handle. One official example was a 1950s-style family illustration showing a family standing in front of a massive cylindrical space habitat filled with green landscapes, lakes, and futuristic buildings.
The company’s product page also showed a Yichang travel layout that placed destination details, attraction information, images, and text on one page as a full travel guide. That requires the model to handle several images, layered text, and page composition at the same time rather than succeeding on only one isolated element.
Another example was a 3x3 grid of medieval modernist posters, each carrying different geometric forms and text, including “Create,” “Grow together,” and “Choose kindness.”
There was also an impressionist street scene of San Francisco, with a road running downhill between colorful houses toward the bay and the Golden Gate Bridge.
OpenAI showed several other examples as well: a cream-and-gold Lake Como wedding invitation, an inverted futuristic city, eight vintage U.S. national park stamps, a solar-flare educational slide deck, a blue-and-gold Earth mosaic, a ChatGPT sticker poster, and a futuristic city in the rain at night.
Taken together, those samples span illustration, posters, invitations, infographics, educational visuals, and promotional imagery. OpenAI’s message was fairly clear: when a prompt defines layout, text, style, and specific visual elements at the same time, Images 2.5 is meant to follow that instruction set more completely.
Fashion entrepreneur Yana Welinder said after testing that Images 2.5 showed a clear improvement in fashion design. In her view, older models could preserve the reference design but sometimes produced a flatter result, while 2.5 delivered a fuller visual outcome.
Not everyone reached the same conclusion. AI and software engineer Mark Kretschmann ran tests focused on noise and artifacts in forest scenes.
He said version 2.5 includes real improvements but did not deliver stable results in his testing. He later compared Images 2.5 with Images 2.0 and said that, in several of his cases, 2.0 looked better in terms of realism.
For now, the safer reading is that Images 2.5 improves edit control and execution on complex instructions, while results still vary depending on the type of image being generated.
Sketch, templates, and image markup extend the workflow
Beyond the model itself, OpenAI added several new ways to work inside ChatGPT.
Sketch is the most direct. Users can draw a rough image in ChatGPT and ask the model to turn that draft into a final picture. A room can start as a basic layout. A clothing concept can begin as a rough outline. Even a simple hand-drawn composition can be used as the starting point, with style and other requirements added later.
Typing “@Sketch” calls the feature. In practice, that reduces dependence on text prompts alone. Some compositions are difficult to describe with words, especially object placement, proportion, and overall shape. Now the user can draw first and let the model complete the rest.
X user @fquolodasha spoke highly of GPT-Image2.5’s hand-drawn input capability after testing it, saying the result was “kind of amazing” and noting how quickly OpenAI has been shipping updates.
Templates target a different friction point. ChatGPT now includes common creative formats such as Poster and Merch. Users choose a template first, then fill in the message, design elements, and visual style instead of starting from a blank canvas every time.
Image sharing now includes a prompt option as well. Users can share the prompt used to create an image together with the image itself. Others can then replace the source photo and specific details with their own and continue from that setup.
OpenAI’s example was an 1980s-style portrait featuring curly hair, a colorful jacket, a gold chain, neon lighting, and a music player. Another user could keep that concept and swap in a different personal photo.
Two API versions: Flare and Sunburst
Images 2.5 is also available through the API. OpenAI introduced GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
Flare is the default option, designed around speed, quality, and editing ability. OpenAI said it produces higher quality than GPT-Image-2 while cutting latency by 50%, making it suitable for social content, product experiences, visual search, rapid prototyping, and large-scale image generation.
Sunburst is aimed at work that needs finer control, such as formal ad assets and high-quality product imagery. It allows longer generation times in exchange for greater editing precision.
Early user feedback published by OpenAI also centered on edit control.
Axultan Alimkulov, head of product at Higgsfield AI, said Flare does a good job understanding what should remain unchanged. For film, UGC, and advertising teams, he said, that means a single edit can preserve the original character, composition, and visual identity as much as possible.
Serial entrepreneur @gkxspace looked at the update through a commercial image-workflow lens. In his view, one of the biggest limitations in AI image generation had been that the first image might be strong, but building a stable second or third image from it was difficult. He said the changes in sequential editing, speed, and reference-image fidelity make scenarios such as e-commerce outfit swaps, brand asset extensions, and serialized illustration more practical.
Images 2.5 is now available to ChatGPT, ChatGPT Work, and Codex users, while Flare and Sunburst are open in the API as well. OpenAI said it is also keeping prompt and image safety checks, C2PA metadata, and invisible watermarking to help identify images generated with its tools.
Based on OpenAI’s own examples and the user tests now in public, the update has a tight focus: make post-generation edits easier to control, keep reference images steadier across repeated use, and add Sketch, templates, and image markup to the workflow.
As for realism, fine details, and image quality after many rounds of editing, the current tests do not point in one direction. More real-world use will be needed before that part is settled.