Can GPT Image 2 Edit Uploaded Images?

OpenAI’s broader image reference says a model can generate a new image from a prompt and/or an input image, which supports input-image workflows rather than text-only generation.^[18] For GPT Image 2 specifically, Replicate describes the model as able to create images from text or edit existing images, and fal.ai exposes an openai/gpt-image-2/edit endpoint whose example request includes a prompt plus image_urls.^[5][7]

What OpenAI’s docs establish

The official OpenAI material in the checked sources is strongest on the workflow category: image generation and image editing are separate documented operations.^[13][15][17] One OpenAI edit-reference snippet also refers to behavior returned by default for the GPT image models, which connects the edit method to the GPT image model family, although that snippet does not by itself spell out GPT Image 2’s full capability list.^[14]

That distinction matters because an edit workflow starts from an existing visual input and produces a new image, while a generation workflow starts from a text prompt alone.^[17][18] In practical terms, GPT Image 2 should not be described as only a new-image generator when the available GPT Image 2 integration pages explicitly document editing existing images.^[5][7]

Uploaded file, image URL, or input image?

The safest wording is input image. fal.ai’s GPT Image 2 edit example uses image_urls, so that integration clearly accepts image URLs as image inputs.^[7] OpenAI’s own reference wording is broader and says prompt and/or input image, without the snippet here showing every native OpenAI transport detail for GPT Image 2.^[18]

That means developers should avoid assuming that a provider wrapper parameter such as image_urls is identical to the direct OpenAI API schema. The checked OpenAI GPT Image 2 model-page snippet does not expose the full request schema, input limits, or account-specific availability, so those details should be verified in the current OpenAI model page and image-edit reference before production use.^[1][15]

Masked editing is supported, but masks are not a hard guarantee

OpenAI’s GPT Image cookbook describes an edit workflow where a mask can be supplied if you do not want the model to change a specific part of the input image.^[19] The same cookbook note cautions that the model might still edit some parts inside the mask and recommends using an image segmentation model if an exact mask is required.^[19]

So masks are useful for guiding edits, but the provided documentation does not support treating them as pixel-perfect boundaries.^[19]

Practical checklist before building

• Use an edit workflow when the task starts from an existing image; OpenAI documents edits as modifying existing images, separate from generations from scratch.^[17]
• Confirm the current direct OpenAI endpoint, request fields, and model availability instead of relying only on a third-party wrapper example.^[1][7][15]
• Treat image_urls as confirmed for fal.ai’s GPT Image 2 edit integration, not automatically as the universal OpenAI-native parameter name.^[7][15]
• If using masks, plan for approximate protection rather than guaranteed unchanged pixels.^[19]

Bottom line

Yes: GPT Image 2 can edit supplied input images; it is not limited to generating brand-new images from text. The strongest general support comes from OpenAI’s documented edit and input-image workflows, while the strongest GPT Image 2-specific examples come from Replicate and fal.ai; developers still need to verify the current native OpenAI schema and limits before shipping.^{[1][5][7][15][17][18]}