Reference image editing

Image-to-image AI: edit pictures with a reference and a prompt

Start with an image you have permission to use. Describe what should stay and what should change, then use the image generator to choose an available model and create a variation.

Open image generator

Opens the workspace. Upload your reference there and review the available settings and credit cost before generating.

From reference to a new image in four steps

  1. 1. Choose a clear reference

    Use an original with a clearly visible subject. Avoid heavy compression and distracting overlays. Only upload images you have the right to use.

  2. 2. Separate what stays from what changes

    Name the subject, composition and details to preserve. Then describe one change, such as a new background, lighting or visual style. The prompts below are starting points, not verified output examples.

  3. 3. Choose a model and settings in the workspace

    Open the image generator, upload your reference and paste your prompt. Check which settings the selected model supports, such as aspect ratio or output size, and review the displayed credit cost.

  4. 4. Inspect the generated variation

    Compare faces, hands, text, product details and framing with the original. If an important detail changed, make the instruction more specific and try another variation. An exact match is not guaranteed.

Three complete prompts to try

Replace a product background

Use the uploaded product photo as the reference. Keep the product shape, proportions, material, label and camera angle unchanged. Replace only the background with a warm, off-white studio backdrop. Add a soft contact shadow underneath the product and keep the original crop. Do not add props, new text, logos or extra objects. Preserve the visible product details as closely as possible.

Change a portrait background

Use the uploaded portrait as the reference. Preserve the person’s facial features, expression, hairstyle, clothing, pose and framing. Replace the background with a softly blurred garden in natural daylight. Match the background lighting to the existing light on the face. Do not retouch the skin, change the person’s age or add accessories. Keep the person clearly separated from the background.

Adjust the style while keeping the composition

Use the uploaded image as the reference. Keep the placement, scale and silhouette of the main subjects, the camera perspective and the crop. Render the scene as a watercolor illustration with visible paper texture, soft pigment transitions and a muted color palette. Preserve the original subject relationships and recognizable details. Do not add or remove objects, text or borders.

Input limits and realistic expectations

  • For the default image upload policy, use JPEG, PNG, WebP or GIF files up to 8 MiB each. GPT Image 2 accepts JPEG, PNG and WebP up to 30 MiB each. The selected model can have additional reference and setting limits; check the workspace before uploading.
  • Outputs can change details you asked to preserve. Review every result before using it, especially faces, text, labels and product claims. Processing time and results vary by model and request.
  • Availability, account requirements and credit costs are shown in the workspace. This guide does not promise free generations or a particular turnaround time.

Frequently asked questions

How is image-to-image different from text-to-image?

Image-to-image uses an uploaded reference alongside your written instructions. Text-to-image starts from a written description without that reference. The reference guides the output but does not guarantee that every detail stays unchanged.

Does copying a prompt also upload my image?

No. Copying only places the prompt text on your clipboard. Open the image generator, upload your own reference and paste the prompt before reviewing settings and starting generation.

Can I use the result commercially?

Check your rights to the input image, the applicable model conditions and our Terms of Service before using a result. A generated image is not a guarantee of permission to use someone’s likeness, brand or copyrighted material.

Why did the model change a detail I wanted to keep?

Generative editing can reinterpret the reference. State the details to preserve explicitly, reduce the number of simultaneous changes and compare the output with the original. Exact preservation is not guaranteed.

Read the Terms of Service

Continue learning

Open image generator