GPT Image 2.5 Sunburst (Custom)

Overview

GPT Image 2.5 Sunburst (Custom) is an OpenAI image model setup for premium image generation and edits. You provide an instruction plus optional reference images. The model produces new images or rewrites parts of an existing image while keeping key details stable. It’s useful when you need tight control across iterative revisions, like product or campaign visuals.

What you can build

• Photoreal product shots with consistent lighting, materials, and branding across variants
• Architectural visualization from sketches, with strict layout and proportion adherence
• Marketing creatives that need repeated “change one thing, keep everything else” edits
• Background removal and asset creation with real transparency for design pipelines
• Targeted retouching with a mask, like swapping text, logos, or colors in one region
• Multi-reference style and identity guidance using several images in a single job

Inputs

• One required written instruction that describes the image or the edit. It can be very long, up to 32,000 characters.
• Optional reference images to guide the result or to perform an edit. You can attach up to 16 images per request.
• Optional mask image for targeted edits. This is used only when you also provide an input image. The mask should match the first input image’s dimensions.
• Your requested output width and height in pixels. Each side must be between 480 and 3840 pixels. The long edge can’t exceed 3 times the short edge.
• A quality choice that trades off detail versus cost. Options range from low through max.
• An optional background choice: automatic, fully opaque, or transparent. Transparent output requires PNG or WebP.
• An optional output file format choice: PNG, JPEG, or WebP.
• An optional compression level from 0 to 100 for JPEG or WebP outputs.
• How many images you want back from one request. You can request 1 to 10 images.
• An optional moderation setting that controls how strict content filtering is.

Outputs

• One or more generated images in your selected format (PNG, JPEG, or WebP).
• Images at your requested pixel dimensions, with automatic snapping and resizing when the size isn’t valid.
• Support for custom resolutions under combined constraints: - Total pixels between 655,360 and 8,294,400
  • Aspect ratio between 1:3 and 3:1
  • Dimensions are adjusted to align with model requirements, then resized to your requested size
• If you request a transparent background and a compatible format, the output includes an alpha channel.

Limitations

• This model returns images, not text. Don’t expect captions, labels, or JSON in the response.
• Very small sizes won’t work. The request must meet the minimum total pixel requirement.
• Masks must be well-formed and aligned. A mismatched mask size or unclear mask edges can cause artifacts.
• With multiple reference images, the model may blend features unless your instruction assigns each image a clear role.
• Low-quality inputs can reduce edit fidelity. Heavy JPEG compression, motion blur, tiny faces, or noisy scans can degrade results.
• If you request transparency, you must choose PNG or WebP output. JPEG can’t carry alpha.

Safety & compliance

• Follow OpenAI’s content and usage policies when generating or editing images.
• The system may block or refuse requests for disallowed content, even if your prompt is detailed.
• Don’t upload images you don’t have rights to edit or redistribute.
• Some outputs may include provenance signals, like metadata or watermarking, depending on the distribution channel.