On Tuesday, September 9th, OpenAI officially unveiled ChatGPT Images 2.5, a next-generation image generation model that emphasizes faster output, higher fidelity, and more precise multi-turn editing. Compared to the previous Images 2.0 version, the company says generation latency has been reduced by as much as 50%, while the model now does a better job at preserving the likeness of reference subjects like people and pets, maintaining detail and composition consistency even after multiple rounds of edits.
This update is less about simply producing a more attractive picture and more about lowering the skill barrier for AI-assisted visual creation. OpenAI has introduced a new feature called Sketch, which lets users draw rough drafts directly in ChatGPT for the AI to turn into full images. A new template system helps streamline the creation of common visuals like posters and product shots. Additionally, users can now comment directly on images, circling specific areas they want changed, so the AI modifies only those sections without requiring users to continually rewrite their prompts. OpenAI has also launched two API models for developers: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare balances speed, quality, and editing functionality with up to 50% lower latency than GPT-Image-2, while Sunburst is geared toward professional workflows that demand finer control, allowing longer processing times for greater precision.
The company’s primary focus for this release is fidelity. Images 2.5 better utilizes reference photos to transpose familiar faces and pets into different scenes, visual styles, and compositions, while enhancing the retention of lighting, texture, and distinctive characteristics. For commercial users creating ongoing content series, this means a person, product, or brand visual is far less likely to morph into something inconsistent after repeated edits. This addresses a long-standing pain point where a model might produce an excellent first result, but any localized change would distort facial features or the overall layout. OpenAI has stressed this multi-turn editing coherence. In extended conversations, Images 2.5 can now retain the history of prior changes, building upon the latest version rather than generating an entirely new, similar-looking image each time.
In early testing, media outlet Axios observed similar improvements. After testing with people, pet photos, tattoo designs, and logos, they found Images 2.5 markedly better at preserving features and executing localized tweaks. For example, when converting a photograph of a cat into film noir, stained glass, and Renaissance portrait styles, the model still captured the original feline's distinct look. Beyond the quality enhancements, the true novelty lies in how editing works now. Users can place notes directly on an image, directing the AI to "delete this area" or "change this part to blue," prompting localized edits that leave the rest of the image intact. Axios’s tests found this especially effective for iterative projects like tattoo designs, where users upload an existing photo, request new elements, and circle regions to adjust color and detail. TechRadar also praised this feature as one of the most practical improvements, noting that users can now point and click to fix elements instead of constantly describing things like "brighten this" or "remove the person in the background." The editor also includes functions for background removal, object erasing, and resizing, with version history allowing users to backtrack. This signals a shift for AI photo editing from being language-driven to visually interactive, where users can simply indicate an object onscreen without having to describe its exact position.
Another significant change is the new Sketch feature. Users can sketch a rough outline directly in ChatGPT—be it a room layout, a clothing silhouette, or a simple object—and the AI will generate a complete image based on that prompt. OpenAI notes you can access this by typing "@Sketch" in ChatGPT. TechRadar’s testing found that Sketch doesn't require drawing ability; even simple outlines drawn with a mouse or finger serve as sufficient visual references. After generation, users can continue to refine the image in edit mode. Axios singled out Sketch as a critical addition, and OpenAI’s head of image products, Adele Li, explained that tools like Sketch and templates are meant to get users more involved in the creative process, rather than making them passive observers. She stated OpenAI's goal is to break away from a so-called "singular AI aesthetic," remarking: "I don't want to look at something in the real world and instantly tell it was made by ChatGPT. I want people to generate and express their own personalities." This reflects a changing philosophy for OpenAI: the AI is not just making images for you, but serving as a creative tool that gives users more control.
The template feature is also aimed at lowering the barrier to entry. Instead of facing a blank canvas, users can start with common formats like posters or product photos and then layer in text, design elements, and stylistic choices. TechRadar tested the product template workflow, uploading a product photo and selecting from options like clean studio shoot, lifestyle editing styles, or specific scenes. They suggested this makes the tool far more approachable for users who know what they want but lack the design skills to execute it. OpenAI is also tailoring Images 2.5 for business use cases, beyond individual creativity, optimizing for marketing, retail, and media production including brand assets, website visuals, product images, and serialized content. OpenAI reports that users now generate over three billion images weekly through ChatGPT Images and the GPT-Image API models.
On the developer side, two new models are now available. The GPT-Image-2.5 Flare model is positioned as the default choice for most applications, focusing on speed, quality, and edit capability, boasting a 50% latency reduction over GPT-Image-2 and higher quality outputs for social content, creator apps, product experiences, visual search, and large-scale generation. Meanwhile, GPT-Image-2.5 Sunburst targets a more professional niche, designed for high-end visual workflows like advertising creative, production-grade marketing assets, and detailed product photography, trading speed for a greater degree of control. Early adopters include Higgsfield AI, Adobe, and Manus. Manus reported that in its evaluations, Flare generated images two to four times faster than GPT-Image-2 while also improving transparent background generation.
OpenAI’s move from version 2.0 to 2.5 isn’t just about generating more photorealistic content; it’s a concentrated effort to enhance control, revision, and ongoing creation. TechRadar, after 24 hours of continuous testing, concluded that while this update may not be a generational leap like "Images 3.0," the improvements are substantial enough to change daily usage habits, predicting that localized editing, Sketch, and templates will become indispensable tools. Axios also observed that as the model’s command over character traits and incremental edits strengthens, the line between AI-generated imagery and traditional design software is blurring. However, Axios reminded that significant ethical debates persist around AI art, specifically concerning whether creators have consented to and been compensated for the use of their artwork in training data. By introducing Sketch, templates, and visual editing, OpenAI is doubling down on pulling users back into the art-making loop. If early AI generation solved "you describe it, I’ll paint it," Images 2.5 is tackling "you show me what’s wrong, I’ll fix it." That distinction is what makes this release noteworthy: AI imagery is maturing from a one-shot "picture generator" into an interactive design partner that supports continuous dialogue, iterative refinement, and genuine creative workflows.
Comments