GPT Image 2 Guide to Reference Images and Precise Editing

GPT Image 2 is OpenAI's hosted model for image generation and editing, accepting text and image inputs and returning images. Developers can call it directly through the Images API using gpt-image-2, or select its documented dated snapshot. It has no open weights for local inference; the official model page describes service access rather than a downloadable checkpoint. GPT Image 2 documentation

What are the key facts about GPT Image 2?

How does GPT Image 2 use reference images for editing?

Reference-image editing lets an application supply visual material alongside written instructions. OpenAI's guide documents edit requests with one or more image inputs and examples combining references into a new composition. That is a useful starting point when appearance has already been established and the task is to revise the scene. OpenAI image generation guide

GPT Image 2 automatically processes image inputs at high fidelity. Its API does not allow the caller to change that behavior with input_fidelity. For implementation, omit that parameter and concentrate on supplying suitable references and a precise edit instruction. OpenAI image generation guide

Write an edit brief with two parts: the intended change and the details to inspect afterward. For a product photograph, the change might be the background; the review list might include the object outline, surface color, printed label, and visible accessories. This makes the result assessable without treating a preservation request as a guarantee.

Use the reference for information it can actually show. If a detail is too small to inspect in the input, decide whether to supply a clearer reference or relax that requirement before generation. Avoid evaluating the output against an undocumented expectation introduced only after seeing the image.

What are GPT Image 2's editing and output limits?

OpenAI warns that GPT Image models can struggle with recurring visual details and exact composition or lettering. High-fidelity input processing does not remove those limitations. Check the returned image against the reference wherever the task depends on identity, text, or placement. OpenAI image generation guide

Output size has concrete constraints. GPT Image 2 supports flexible dimensions, but each edge must be a multiple of 16, the longest edge cannot exceed 3840 pixels, and the aspect ratio cannot exceed 3:1. Total pixel count must fall between 655,360 and 8,294,400; outputs above 3,686,400 pixels are described as experimental. OpenAI image generation guide

Transparency is also version-sensitive. OpenAI's August 20, 2026 changelog announces transparent backgrounds in preview for GPT Image 2 and its dated snapshot. It specifies PNG or WebP output for that feature, with JPEG excluded. Treat preview support as a feature to test on the exact output you need. OpenAI changelog

Image inputs affect cost. OpenAI's standard pricing table lists GPT Image 2 at $8 per million image input tokens and $30 per million image output tokens, with text input at $5 per million tokens. Those are token rates, not a fixed price for every edit. OpenAI pricing

For a trial, retain the original image and review the result at the size where it will be used. Inspect small lettering at full resolution as well as the overall composition. Record any unintended change as a failed requirement even if the new background is attractive.

How do you edit an image with the GPT Image 2 API?

Configure authorized API access, then choose a reference image that you are entitled to upload and edit. Use the direct Images API when you want to select the image model explicitly. The following request follows OpenAI's documented multipart editing pattern with an original example instruction. OpenAI image generation guide

curl https://api.openai.com/v1/images/edits \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -F 'model=gpt-image-2' \
  -F 'image[][email protected]' \
  -F 'prompt=Place this ceramic mug on a pale wooden table beside a window. Preserve the mug shape, blue glaze, and handle position. Add no lettering.'

Use an existing local file named reference.png, or change the file argument to your actual input. The API returns image data in b64_json; decode the returned data and save the result for inspection. The request omits optional output settings. OpenAI image generation guide

For a reproducible editing trial:

1. Save the original reference separately from generated output.
2. Write the intended change and a short list of details that must remain acceptable.
3. Submit the edit with an explicitly chosen model identifier.
4. Save the result, input instruction, and response metadata relevant to your evaluation.
5. Compare the reference and output side by side.
6. Revise only the part of the brief associated with the observed error, then repeat if useful.

If you need to hold the model version constant, OpenAI documents gpt-image-2-2026-04-21 as a snapshot identifier. Record whether your trial used that snapshot or the general model alias. Version tracking makes later output comparisons easier to interpret. GPT Image 2 documentation

For evaluation purposes, change one important requirement at a time. Try the background edit before adding a new camera angle, extra objects, and a rewritten label. Combining every change in one request makes it harder to identify which instruction needs refinement.

The GPT Image API integration guide covers request handling and cost evaluation. For organizing the wider creation process, see the ComfyUI guide; verify the exact integration before assuming any workflow supports this hosted model.

How does GPT Image 2 compare with GPT Image 1 for editing?

For a migration test, start with the same source image and approved edit brief. Judge whether the output meets the requirement, then compare observed usage and completion behavior. A different-looking image is not automatically better or worse; its value depends on the details the job requires.

Keep a few demanding references alongside ordinary ones. Include small lettering, an unusual silhouette, or a composition with narrow margins when those occur in your actual work. Decide in advance which deviations require another edit and which can be accepted.

What else should you know about GPT Image 2?

Is GPT Image 2 an official OpenAI model?

GPT Image 2 is an official OpenAI model released on April 21, 2026. Its model page documents the API identifier gpt-image-2 and the snapshot gpt-image-2-2026-04-21. OpenAI changelog GPT Image 2 documentation

Can I download GPT Image 2 weights?

OpenAI provides GPT Image 2 as a hosted image generation and editing service, without open weights. Use its authenticated API access path for the documented integration. GPT Image 2 documentation OpenAI image generation guide

Should I set input_fidelity to high?

For GPT Image 2, omit input_fidelity. OpenAI says the model processes image inputs at high fidelity automatically and does not permit changing that setting. OpenAI image generation guide

Can GPT Image 2 produce a transparent background?

GPT Image 2 supports transparent backgrounds in preview, announced on August 20, 2026. Request background=transparent with PNG or WebP output, then inspect the resulting edges in your intended composition. OpenAI changelog

Sources

• GPT Image 2 model documentation
• OpenAI image generation and editing guide
• OpenAI API changelog
• OpenAI API pricing
• OpenAI API deprecations

Related guides on PromptZone

• Best SDXL Models in 2026
• ComfyUI 2026: The Complete Guide
• AI Model Releases Timeline