GPT Image 2.6

Positioning Overview

gpt-image-2.6-flare is an image-to-image model aimed at developers who need to modify, extend, and re-create existing assets. It takes source images as input and prompts as instructions to complete edits, accepts up to 16 reference images per call, and supports mask-based local repainting, annotation prompts, and 1K to 4K resolution output. It is suitable for embedding image editing capabilities into automated workflows such as content production, e-commerce assets, and marketing imagery.

Core Highlights

• Accepts up to 16 source images per call — multiple reference images can serve as editing bases at the same time, reducing batched calls and result stitching.
• Two local control methods: masks and annotations — masks specify the repaint region, while annotated images are used to circle prompts and are not retained in the output; the two cannot be used together.
• Quality and complexity adjusted separately — rendering quality and planning complexity each have their own levels, allowing trade-offs based on task difficulty without upgrading or downgrading everything at once.
• Resolution and aspect ratio can be specified per output target — supports 1K, 2K, 4K, and a variety of aspect ratios, and can also follow the source image or let the model decide automatically.

Capability Description

Instruction-based editing from source images: describe the intended modification in a prompt and pair it with 1 to 16 jpeg, png, or webp source images to complete the edit, suitable for targeted remodeling on top of an original image rather than generating from scratch.

Mask-based local repainting: pass in an RGBA PNG mask of the same size as the first source image; transparent pixels mark the edit region. Areas outside the mask may also change slightly.

Annotation prompts to locate modifications: use a transparent PNG with annotations such as circles as an auxiliary prompt to guide the model's attention to the target region; the annotations themselves do not appear in the output.

Output format and background control: choose png, jpeg, or webp output; jpeg and webp support compression levels from 0 to 100. When a transparent background is needed, the output format is limited to png or webp.

Separate levels for quality, complexity, and content moderation: rendering quality has five levels and planning complexity has three, allowing separate adaptation to fine retouching and to tasks with dense text and complex layouts; content moderation can be set to automatic or low.

Use Cases

• Design teams iterating on existing assets over multiple rounds — input an existing visual draft and modification notes, adjust composition and details round by round, and produce image assets ready for review.
• Asset re-creation for e-commerce and brand teams — use product images or key visuals as source images to replace backgrounds, extend scenes, or correct details, obtaining product and campaign images ready to publish.
• Compositing and local retouching workflows — use masks or annotations to locate the areas to change, adjust only the specified parts, and output graphics for layout compositing such as posters and detail pages.
• High-resolution delivery needs — directly specify 2K or 4K resolution and the target aspect ratio, reducing post-processing upscaling and cropping steps.

Limitations and Selection Advice

This is an image-to-image editing model and does not have text-to-image capability; to generate images from scratch, use the text-to-image route in the same series. Local control relies on masks or annotations, and areas outside the mask may also change slightly. Transparent background output is constrained by format and must be paired with png or webp; compression levels apply only to jpeg and webp. When details must stay consistent across multiple rounds of editing, the Sunburst route in the same series can be used.