GPT Image 2: FAQ

1. What is GPT Image 2?

GPT Image 2 is the current flagship OpenAI image model exposed on Luma. It generates images from text prompts and edits existing images via a source image plus optional references. It delivers higher fidelity than GPT Image 1.5 on complex prompts and multi-image compositions, and it preserves the context of input images including styles, composition, and other details. It is slower and more resource-intensive than other image models, and its high quality tier is the most expensive image-generation option on the platform.

2. What resolutions are available?

Three resolution tiers: 1K, 2K, and 4K. The default is 2K. Higher resolutions cost more. Successive edits degrade quality, so run an edit chain at the final resolution rather than stepping up later. There is no image upscale tool to recover detail after the fact.

3. What aspect ratios are supported?

Ten aspect ratios: 3:1, 21:9, 16:9, 3:2, 4:3, 1:1, 3:4, 2:3, 9:16, and 1:3. The default on create is 16:9. The model snaps the output to the nearest supported ratio with no deviation cap, making it safe for any source image shape. 3:1 and 1:3 sit at OpenAI's max ratio boundary. On the modify tool there is no aspect-ratio parameter; the output follows the source image's framing.

4. How do reference images work?

Create: up to 16 reference images can be attached. References can supply style, character, pose, or composition guidance. When using two or more references, label each image's role explicitly in the prompt (e.g. "Use image 1 as character reference. Use image 2 as background/style reference.").

Modify: one source image is required, plus up to 15 additional references (16 total including the source). The same role-labelling approach applies. GPT Image 2 is the strongest tool for literal face copy-paste on non-celebrity or unfamiliar faces, with minimal identity drift. For pose or expression variations on a recognizable character, a reference-following model such as Nano Banana Pro may be preferable.

5. What quality and format options are available?

Quality has three levels: low (fast), medium (balanced, default), and high (best quality). High quality on GPT Image 2 is more expensive than the high tier of any other image model.

Output format has three options: png (lossless, supports transparency), jpeg (smaller file), and webp (modern compression). The default is png.

6. Does it support transparency?

Yes. The background parameter has three values: auto (model decides, default), transparent (RGBA cutout for compositing, available with png or webp), and opaque (solid background).

7. Create vs Modify: what is the difference?

The create tool (create_image_gpt_image_2) generates an image from a text prompt, with up to 16 optional reference images and a full set of aspect-ratio, resolution, quality, format, and background controls.

The modify tool (modify_image_gpt_image_2) edits an existing image. It requires a source image and accepts up to 15 additional references (16 total). It shares the same resolution, quality, format, and background controls, but it has no aspect-ratio parameter: the output inherits the source image's framing. Use modify for targeted edits, face swaps, multi-reference compositions, and complex composition changes on an existing image.

At a glance

GPT Image 2 vs GPT Image 1.5

• Aspect ratios: GPT Image 2 supports 10 ratios (3:1 through 1:3). GPT Image 1.5 supports only 3 (1:1, 3:2, and 2:3), and will crop or resize outputs that do not match.

• References: GPT Image 2 accepts up to 16 references (create) or 16 total including source (modify). GPT Image 1.5 accepts up to 4.

• Resolution: GPT Image 2 exposes 1K, 2K, and 4K. GPT Image 1.5 has no resolution parameter.

• Fidelity: GPT Image 2 delivers higher fidelity on complex prompts and multi-image compositions.

• Cost: GPT Image 2's high quality tier is more expensive than GPT Image 1.5's high tier.