ChatGPT Images 2.5 · Official Guide to Sketch, Templates, Sunburst and Flare

Quick Summary

• Credit: The four embedded videos, demos, and release data belong to OpenAI. This article organizes the official information and adds editorial recommendations and prompts from JQ VISIONS.
• OpenAI announced ChatGPT Images 2.5 on September 8, 2026 as an update with sharper details, faster generation, more accurate editing, and new ways to create and share.
• The company claims that generation latency has been reduced by up to 50% compared to Images 2.0, while improving lighting, texture, subject recognition and preservation of references.
• Localized editing seeks to change only the requested element, maintaining the subject, composition and brand treatment. Multi-turn consistency allows you to accumulate changes with less degradation throughout the conversation.
• Sketch transforms a drawing made directly in ChatGPT into a visual reference. It is activated with @Sketch and can be accompanied by a description of style, materials and details.
• Templates offer starting points for popular formats, including posters and merch. The user adds messaging, design elements, and visual direction rather than starting with an empty prompt.
• Comments can be placed directly on the image to guide local corrections. When sharing an image, you can also include the prompt for others to adapt the idea.
• In the API, GPT-Image-2.5 Flare is the default choice for most applications and favors speed. GPT-Image-2.5 Sunburst is intended for premium work that benefits from greater precision between edits.
• Both API models accept text and image as input and support transparent backgrounds, image generation and editing, and quality levels from low to max, as well as auto.
• Images 2.5 is available to ChatGPT, ChatGPT Work and Codex users on desktop, mobile and web. Flare and Sunburst are available in the API.

The official launch in less than a minute

The first video summarizes the ChatGPT Images 2.5 proposal on four fronts: faster generation, greater visual fidelity, better preservation of details during multiple edits and comments placed directly on the image.

The official page adds scale to the context. According to OpenAI, more than three billion images are created weekly between ChatGPT Images and the GPT-Image models in the API. Images 2.5 was designed to improve the entire cycle, not just the first generation.

There are two layers to the release. ChatGPT Images 2.5 is the product experience with chat, Sketch, templates, comments and sharing. For developers, GPT-Image-2.5 Flare and Sunburst are selectable models in the Image API or in the Responses API image generation tool.

The distinction avoids common confusion. Visible functionality in ChatGPT, such as Sketch, is not automatically an API parameter. Likewise, explicitly choosing Flare or Sunburst is a documented decision for the API.

More speed without losing reference

OpenAI claims to have reduced generation latency by up to 50% compared to Images 2.0. The expression is even important: the real gain depends on the order, quality, size, edition, load and product used.

Enhanced fidelity seeks to keep people, places, products and memories recognizable when they are placed in other environments, compositions or styles. The company highlights more natural lighting, richer textures and greater preservation of distinctive features.

For production, the right metric isn't just time per image. Measure how many images make it to approval, how many rounds are needed, and how much manual work remains. A generation twice as fast may not reduce costs if it requires correcting identity, text or product.

Use a clean and sufficiently large reference. Identify the role of each image, list the mandatory details and clearly separate the aspects that can be reinterpreted from those that function as invariants.

Localized editing and comments on the image

Images 2.5 has been optimized to change only what is requested. OpenAI exemplifies the exchange of a product, background or block of text without reconstructing the surrounding subject, composition and brand treatment.

In ChatGPT, comments add a spatial layer. You can click or mark a region and associate a correction with it, which is useful when there are several similar objects or when words like left and right would be ambiguous.

Visual marking does not replace instruction. Write the change completely and define what should remain. For example: replace only the marked packaging, keep hands, reflections, shadow, perspective, background and all other texts.

Make one change per round when the asset is important. Then compare the new version with the previous one using an overlay. This way you find drift in scale, face, negative space or visual identity that goes unnoticed on quick observation.

Multi-turn consistency: edit without starting over

In a long conversation, each image may depend on decisions made several messages earlier. OpenAI claims that Images 2.5 better preserves these decisions and degrades quality less as new changes are applied.

This is relevant for object rotation, campaign variations, interior spaces, characters and infographics. Instead of rebuilding everything, the team can freeze a base and iterate on specific details.

Still, consistency should not be confused with parametric control. One generation continues to be probabilistic. Save an approved version, maintain an inventory of invariants, and return to baseline when a sequence of edits begins to accumulate unwanted changes.

A good production history includes image, references, prompt, model, quality, date and purpose of each edition. Without this record, a long conversation makes it difficult to identify at which step an error appeared.

Sketch: draw before describing

Sketch was created for ideas whose composition is easier to draw than explain. Just write @Sketch in ChatGPT, draw directly in the interface, and use that drawing as a guide for the final image.

OpenAI suggests cases such as the layout of a room, the outline of a piece of clothing or a doodle. You can add a description of the style, materials, lighting and other details that the scribble doesn't communicate.

The sketch should indicate relationships and not beauty. Use simple shapes for position, scale, direction, horizon, and empty space. Then, in the text, say what is rigid and what the model can reinterpret.

For accurate layouts, place short labels on the drawing and repeat them in the instruction. Confirm in the result that the relationship between objects was preserved, because Sketch guides the generation but does not work like CAD.

Templates and prompt sharing

Templates solve the blank page problem. The user chooses a popular format, such as Poster or Merch, and adds messaging, design elements, and visual direction to adapt the starting point.

A template is not a ready-made identity. To avoid generic results, provide audience, purpose, hierarchy, palette, tone, approved text, tag constraints, and final format.

Sharing has also changed. When sharing an image, the user can choose to include the prompt. Another person can reuse the structure with their own photographs and details.

Treat a shared prompt as an editorial template. Remove personal data, private references and customer names, mark replaceable fields and include rights and validation rules.

GPT-Image-2.5 in the API: two models, two rhythms

OpenAI's technical video features GPT-Image-2.5 Sunburst and Flare. Sunburst is the company's most capable model for generating and editing images. The demo highlights sharper detail, natural lighting, rich textures, greater control across multiple edits, and fine instructional following.

Flare is the fastest model in the family. The official page calls it the default choice for most applications and claims that it delivers higher quality images than GPT-Image-2 with 50% less latency. The video describes it as over 50% faster while maintaining equivalent quality.

In the Image API, directly specify gpt-image-2.5-flare or gpt-image-2.5-sunburst. In the Responses API, select a supported master model and define the model within the imaging tool.

The template pages indicate text and image input, image output, generation and editing, and low, medium, high, xhigh, max and auto qualities. The choice should be made based on the accuracy value, not just the model name.

Transparent backgrounds for reusable assets

Flare and Sunburst support transparent backgrounds. OpenAI points to applications ranging from infographics to games, where an asset needs to be placed on different interfaces, scenes or materials.

Useful transparency is not just about erasing a flat color. Glass, water, smoke, hair, thin fabric, motion blur and glow require partial alpha. Specify these zones and request an RGBA PNG rather than accepting an image that merely mimics the checkerboard pattern.

Validate the file on white, black and saturated backgrounds. Look for halos, color contamination, jagged edges, holes, and shadows that are stuck to the previous background.

For UI and games, also test the asset at full size. An outline that looks correct at 1024 pixels may look jagged on a 32-pixel icon or create fringing during animation.

When to choose Flare or Sunburst

Choose Flare when you need to explore many directions, generate social content, prototype products, power visual search, or process volume. Lower latency allows you to compare alternatives before investing in completion.

Choose Sunburst when a campaign, product image, or localized edit requires more precision. It is the option recommended by OpenAI for premium workflows, but accepts longer generation times.

An efficient pipeline can combine both. Generate six concepts with Flare, approve one, then recreate or edit the winner with Sunburst. Preserve reference and invariants so that template passing doesn't transform creative direction.

Evaluate with an actual set of requests from the team. Record time, cost, pass rate, number of rounds, loyalty to referrals, and remaining manual work. The best choice may vary between thumbnails, e-commerce, UI, infographics and campaign.

Four workflows that take advantage of the update

Campaign with person or product: approve references, generate directions with Flare, finalize the frame with Sunburst and make localized changes without rebuilding the rest of the composition.

Infographics: research and approve facts first, manage visual hierarchy, validate all text and complete typography and alignments in a deterministic tool.

Merch and posters: start from a template, replace message and visual language, generate mockups and then separate the final art into a transparent asset for production.

UI or game: draw the spatial relationship with Sketch, generate isolated components with a transparent background, maintain visual tokens and test assets at full size within the product.

In all four cases, the most important passage is approval. Set a checkpoint between exploration and master, because a conversation without decision accumulates variations without transforming any of them into a ready asset.

Security, rights and provenance

OpenAI claims that Images 2.5 maintains checks on prompts and images to prevent harmful output. The company also continues to use C2PA metadata and invisible watermarking to help identify content created with its tools.

Technical provenance does not replace authorization. Confirm rights to photos, faces, products, logos, styles, characters and text before generating or publishing.

When the image represents a real person, preserve context and do not place the person in misleading situations. When reporting facts, keep sources out of the picture and human review before distribution.

If you share the prompt, remove personal data, private files, customer names and internal details. Sharing the method is helpful, but the prompt should not turn into a leak of information.

Availability and how to get started

According to OpenAI, ChatGPT Images 2.5 is available to ChatGPT, ChatGPT Work and Codex users on all plans, across desktop, mobile and web. Access may appear gradually during rollout.

You can directly open the ChatGPT Images area, start a normal conversation, select a template or type @Sketch. For editing, upload an authorized image, state the objective and indicate what should not change.

Developers can call Flare or Sunburst through the Image API or the image-generation tool in the Responses API. Official documentation warns that some organizations may need verification before using GPT Image models.

Start with a small, measurable case. Choose a poster, a transparent asset or a product edition, define acceptance criteria and compare the new process with the previous one. The update is only valuable when it reduces rounds and increases the percentage of approved work.