OpenAI launches Images 2.6 with GIFs and faster generation

Solving the consistency problem

One of the primary hurdles for AI image generation has been the inability to maintain a specific subject across different prompts. Users often find that requesting a tiny change to a logo or a character's pose results in a completely different image every time. OpenAI is addressing this with the launch of Images 2.6.

The new model introduces saved references, which allow users to store a specific image of a person, a pet, or a character. Once a reference is saved, the model can reuse that specific visual identity in future requests. This helps ensure that a character looks the same whether they are in a forest, a city, or a bedroom.

By providing a way to keep visual details consistent, OpenAI aims to make the tool more useful for brand work and storytelling. Instead of re-uploading a character every time and hoping for the best, users can now maintain a stable visual thread across multiple scenes.

Speed and animation features

In addition to consistency improvements, Images 2.6 is designed to be significantly faster. OpenAI reports that the new model is about 50% faster than previous versions. For developers using the API, early reports suggest that the speed-focused model can generate an image in as little as six seconds at low quality.

The update also introduces support for GIFs, which are short animated loops. These can be used to bring mascots, product shots, or sketches to life with simple motion. This adds a layer of dynamic content to the existing image generation capabilities available in the tool.

These performance gains are intended to streamline the creative workflow for high-volume users. Whether a user is looking for a quick mockup or a polished animation, the faster generation time reduces the waiting period between prompts. This is particularly relevant given that OpenAI estimates people make about 4 billion images a week across its products.

Tools for professional designers

For professional workflows, the update includes several technical features tailored for developers and designers. Image markup is now available via the API, allowing users to point to a specific spot on an image rather than relying solely on text descriptions. This helps eliminate confusion over which side of an object needs an edit.

The model also supports SVG export, which saves designs as instructions rather than a grid of pixels. This means the resulting shapes stay sharp at any size and can be edited by designers using traditional software. It is a significant step toward making AI-generated assets usable in professional print and web design.

Additionally, a new complexity setting gives users more control over the model's planning process. OpenAI suggests using the low setting for quick edits and the medium or high settings for crowded scenes with many specific constraints. This allows the user to balance the amount of detail the model considers before it begins drawing.

Model variety for different needs

OpenAI is offering two distinct models within the 2.6 release to cater to different user priorities. The GPT-Image-2.6 Flare model is specifically built for speed and is well-suited for layouts like posters. It prioritizes quick results and general composition over minute details.

Conversely, the GPT-Image-2.6 Sunburst model is designed for more detailed work and consistent edits. This model is better for users who need to perform multiple iterations on a single subject while maintaining high fidelity. The two models represent a split between rapid prototyping and high-quality refinement.

These options are now available to all ChatGPT subscription tiers, as well as to developers through the API. By offering different paths for speed versus detail, OpenAI allows users to choose the most efficient tool for their specific project requirements.

Testing the new capabilities

To see if the new consistency features work as intended, users can test the saved references with a single character. The goal is to place that character in five different scenes with minor edits made between each step. If the character in the fifth scene still looks like the one in the first, the consistency pitch holds true.

Users can also experiment with the SVG export to see how well the model handles shapes versus photo-like images. While the source does not specify how well the model handles photorealism, testing a logo first is a recommended way to gauge the output quality. This will show how the model handles clean lines and distinct shapes.

The rollout of Images 2.6 marks a shift toward more controlled and predictable AI generation. By moving away from purely random outputs toward a system that remembers and follows specific instructions, OpenAI is making the tool more viable for complex creative projects.