Edit: Change Existing Images

The Edit tab changes part of an existing image based on a written instruction. The result is placed above the original as a new layer, so the original pixels stay untouched.

How to edit an area

1. 1
   Select the area you want to change. Any Photoshop selection tool works, for example the lasso or the rectangular marquee.
2. 2
   Describe the change in the text field, for example "replace the cup with a red rose".
3. 3
   Click Generate. The result comes back as a new layer with a layer mask in the shape of your selection. Only the selected area appears changed.

Seam automatically extends the area it sends to the model so that it matches a supported image shape. The visible change still stays inside your selection.

Without a selection

• If you do not make a selection, the active layer is used as the editing target. Its whole content is sent.
• If you add images from files, the first image is used as the target instead and the rest become references. Seam warns you once before it does this, so an upload never silently replaces what you thought you had selected.

Sample all layers

This switch sits in the Layer settings row of the panel and controls what the model sees. When it is on (default), the model sees everything that is visible in the target area, across all layers. When it is off, the model sees only the content of the active layer. Turn it off when you want to edit one layer in isolation.

The switch is shown in the Edit and Upscale tabs. Create does not have it, because there is nothing to sample. Background Removal turns it off while it is selected, since it always works on a single image, and your own preference comes back when you switch models.

Models in the Edit tab

The Edit tab shares the Nano Banana models, GPT-Image-2, and Reve 2.1 with the Create tab, and adds five models of its own. In the order the picker shows them:

• Nano Banana 2.1 (default), Nano Banana 2 Lite, Nano Banana Pro, GPT-Image-2, Reve 2.1: Same models and settings as in the Create tab, here applied to your image.
• FLUX.2 Max Edit: Black Forest Labs' top FLUX.2 editing tier. Up to 9 reference images, eight aspect ratios, seed. Each variation is a separate run.
• FLUX.2 LoRA Edit: Can load your own trained models (see "My models"). A LoRA is a small add-on file that gives a model a specific style. Up to 3 reference images, guidance, steps, and acceleration settings.
• Seedream 5.0 Pro Edit: Up to 10 reference images and up to 6 variations per run.
• Grok Imagine 2.0: Up to 3 input images, 1K or 2K output, thirteen aspect ratios, up to 4 variations per run.
• Background Removal: Cuts the subject out of the image. No description needed, the text field disappears. With the "Apply as mask" option (on by default) all original pixels are kept and the background is hidden with a soft layer mask instead of being deleted.