GPT Image 2.6 Sunburst Custom

Overview

GPT Image 2.6 Sunburst Custom is an OpenAI image model for high-detail generation and accurate edits. It renders new images from a written description, or it edits existing images while keeping key details stable across revisions. You can steer results with quality and planning controls, then export in common raster formats or as an animated GIF. This model is useful when you need layout precision, readable text, consistent products or characters, and controlled edit rounds.

What you can build

• Product photo variations that keep packaging and branding consistent
• Poster and flyer concepts with tighter layout control and cleaner typography
• Image retouching workflows like object removal, background changes, and relighting
• Before-and-after design iterations using the previous output as the next edit input
• Brand mascot loops and reaction-style animations as short GIFs
• Guided edits where a circle or arrow overlay shows exactly where to change something

Inputs

• A written description of the image you want, up to 32,000 characters.
• Optional reference or edit images you upload to guide the result, with up to 16 images total.
• For edits, an optional mask image that uses transparency to suggest what may change.
• For markup-guided edits, an optional transparent PNG overlay (same pixel size as the first input image) where you draw circles or arrows.
• The output width and height in pixels, up to 3840 on either edge.
• Size constraints for custom dimensions: both edges should be multiples of 16, the aspect ratio must stay between 1:3 and 3:1, and total pixels must stay within the model’s pixel budget.
• A quality level that trades cost for detail, from low through max.
• A planning level (low, medium, or high) that changes how much the model plans before rendering.
• The number of variations to generate in one request, from 1 to 10.
• An output file format choice: PNG, JPEG, WebP, or GIF.
• An optional background mode: auto, opaque, or transparent. Transparent backgrounds require PNG or WebP.
• Optional compression control for JPEG or WebP outputs, from 0 to 100.
• For GIF output, an optional frame count (2 to 16) and frame duration (20 to 2000 ms per frame).
• An optional moderation strictness choice, which controls how aggressively the system filters disallowed content.

Outputs

The model returns 1 to 10 generated results.

Each result is an image file in your chosen format:

• PNG, JPEG, or WebP for still images
• GIF for a short loop animation

Outputs typically include the image file plus basic file metadata such as format and pixel dimensions. When you request multiple variations, you’ll get multiple separate files.

Recommended settings

• Precise edits (small local changes): use low planning, and raise quality only when you need sharper details.
• Dense layouts (posters, UI mockups, lots of labels): use high planning, then move quality to medium or high for legible small text.
• Multi-round editing (several edits in a row): keep quality at medium or high to reduce visual drift between rounds.
• GIF loops: keep frame count modest first, then increase frames only if motion looks choppy.

Limitations

• GIF output is for new generations only. It is not available when you also provide input images for editing.
• GIF output has tighter feature support than still images. Some edit helpers like masks and custom compression won’t apply.
• If you pick dimensions that are not multiples of 16, the system may snap and rescale during generation.
• High planning can change parts of the image you didn’t ask to change. This is more visible on edits.
• Mask-based edits are guidance, not a hard boundary. The model may not match the mask shape perfectly.
• Low-quality inputs increase failure risk. Blurry photos, heavy JPEG artifacts, scanned documents, and noisy screenshots can cause unintended changes.
• Text rendering is improved, but it is not guaranteed. Always proofread logos, names, and small lettering.

Safety & compliance

This model runs with OpenAI image safety systems. Content may be blocked, altered, or refused when it violates policy.

If you choose a less restrictive moderation mode, filtering may be relaxed, but policy still applies. Do not attempt to generate disallowed content, including sexual content involving minors, instructions for wrongdoing, or hateful content.