GPT Image 2.5 Sunburst (OpenAI)

Overview

GPT Image 2.5 Sunburst is OpenAI’s precision-focused image generation and editing model. It turns your instructions into finished raster images, or it rewrites parts of an existing image while keeping the rest stable. You can anchor edits with up to 16 reference images and optionally constrain changes with a mask. It’s useful when you need clean, repeatable art direction across iterative revisions.

What you can build

• Product photo edits that keep the item and lighting consistent while changing the setting
• Campaign creatives with tight brand control across multiple revisions
• UI mockups, posters, and infographics with dense, readable text
• Character sheets and multi-scene illustrations with stronger identity consistency
• Virtual try-on and wardrobe swaps that preserve face, pose, and background
• Background removal and asset cutouts using true transparency (alpha channel)

Inputs

• A text description of what to generate or what to change. It can be up to 32,000 characters.
• Optional reference images to edit or to guide the output. You can provide 1 to 16 images. Use PNG, WebP, or JPG. Each image must be under 50 MB.
• Optional mask image to limit edits to specific regions. Use a PNG with an alpha channel. The mask must match the first reference image’s exact dimensions.
• Output size selection. You can use common presets or an explicit width × height. If you provide exact dimensions, both sides must be multiples of 16. The longest side can’t exceed 3840 px. Aspect ratio must stay within 1:3 to 3:1. Total pixels must be between 655,360 and 8,294,400. Sizes above 2560 × 1440 are experimental.
• A quality level choice: Auto, Low, Medium, High, Extra High, or Max.
• Background behavior selection: Auto, Opaque, or Transparent.
• Output file format selection: PNG, JPEG, or WebP.
• Optional compression level from 0 to 100 when you choose JPEG or WebP.
• Number of images to generate in one run. You can request 1 to 10 images.
• Optional moderation strictness setting. Choose standard filtering or less restrictive filtering.

Outputs

The model returns one or more image files in your selected format (PNG, JPEG, or WebP). Each result includes the rendered pixel dimensions and file metadata, and you receive one output per requested image. If you request a transparent background with PNG or WebP, the output includes a real alpha channel for compositing.

Recommended settings

• Quick drafts: use Low quality and a standard size like 1024 × 1024, 1536 × 1024, or 1024 × 1536.
• Final creative assets: start at High, then step up to Extra High or Max when fine detail matters.
• Transparent assets: set the background to Transparent and choose PNG or WebP.
• Print-like or crop-heavy work: use larger sizes within the 3840 px edge limit, and compare High vs higher tiers.

Limitations

• Exact typography placement can still fail in dense layouts. You may need a few iterations.
• Element placement in rigid grids can drift. Specify layout constraints clearly.
• Character consistency improves with references, but it can still vary across generations.
• Custom sizes must follow strict constraints (multiples of 16, pixel range, and max edge). Larger-than-2560 × 1440 sizes are experimental.
• Masks must match the base image’s dimensions. Missing alpha in a mask can break targeted edits.
• Low-quality inputs (heavy compression, blur, screenshots with artifacts, or noisy scans) can reduce edit precision and subject preservation.

Safety & compliance

OpenAI filters prompts and generated images under its content policy. You can choose a standard moderation setting or a less restrictive one, but requests can still be blocked by safety checks. OpenAI also uses C2PA metadata and invisible watermarking to help identify AI-generated images. Only upload images you have the rights to edit, and avoid creating misleading or harmful content.