Modify Image

The Modify Image node changes an existing image according to a text instruction. You connect a picture, choose one of more than 20 image-editing models, and describe the change: a new style, a new background or one fixed detail. The node returns a new image and leaves the original untouched.

When to use it

• Restyle a photo or a render, for example into watercolor, anime or a film look.
• Make a targeted change, such as "add a hat to the person" or "make the sky stormy".
• Repaint one area only. Paint a mask, or wire one from Generate Mask or Paint Mask.
• Give an image a new shape and let the model fill the new space.
• Apply the same edit to many images. Wire a List of images into the Image input, and the node runs once for each image.

To enlarge an image without changing it, use Upscale Image. To cut out the subject, use Remove Background. Neither needs a prompt. To separate every element of a picture into its own layer, use Split into Layers.

Quick start

Inputs

The output, Image, is the new picture. It can feed any number of nodes at once.

Settings

The settings panel shows only the settings that the selected model supports. When you switch models, a value the new model does not accept is reset.

Models

Modify Image can run every model in the table. Click a model for its settings, credit prices and prompt tips.

The Provider list also offers GPT Image 2.6 Flare, GPT Image 2.6 Sunburst, MiniMax H3 Image, Seedream 5.5, Nano Banana, GPT Image, Seedream Edit and Flux Fill (Inpaint), which are not in the table.

Which model to choose

• GPT Image 2.6 Flare — the default. It changes what you ask and keeps the rest. Fast edits from up to 16 input images, and each image adds to the price. GPT Image 2.6 Sunburst costs the same and is for finished, detail-critical edits. Both take a mask. See Masks on GPT Image 2.6.
• MiniMax H3 Image — changes or combines up to 16 images in one edit, with readable text inside the picture, up to about 16 MP at 4K. It is priced by size and by each image, the source included. Auto keeps the first image's shape.
• Seedream 5.5 — edits round after round that change what you ask and keep the rest of the picture as it was. It takes up to 10 input images, and each image adds to the price. It keeps natural light, materials and skin, and draws a transparent background in one step.
• GPT Image 2.5 Flare — fast edits from up to 16 source images. GPT Image 2.5 Sunburst is slower and gives the tightest control for demanding retouches.
• Nano Banana — fast, for quick transforms and iteration.
• Nano Banana 2.1 — consistent edits with strong character consistency and text rendering, from up to 10 input images.
• Nano Banana Pro — more detail, for production images.
• Nano Banana Edit — edits from a plain instruction, such as "add a hat to the person".
• Flux Kontext Pro and Flux Kontext Max — context-aware edits that keep the subject's identity. Max is the higher-fidelity tier.
• Ideogram Edit — repaints a masked area.
• Flux Fill (Inpaint) — a model made only for masked edits. It repaints the white area of the mask.
• Seedream 5 Pro — follows detailed instructions and accepts several reference images. Seedream 5 Lite is the budget option. Seedream 5 Flash is fast and low-cost, at the same price at 1K or 2K.
• Flux 2 Max (Safety Tolerance) — accepts up to 8 reference images. Flux 2 Pro (Safety Tolerance) accepts up to 4.

For a comparison of every image model, read Choosing a model.

Change one area with a mask

A mask tells the model which pixels it may repaint. White areas change; black areas stay as they are. The colors mean the same on every mask model.

Instead of painting in the panel, you can wire a mask into the Mask input. Generate Mask finds a subject from a description such as "the red car". Paint Mask keeps a hand-painted mask as its own node, which several nodes can share. When you reopen the painter on a saved mask, it starts from that mask, so you can touch it up instead of starting again.

Masks on GPT Image 2.6

• The rest of the picture is kept exactly as it was. GPT Image 2.6 redraws the whole picture. After the run, the source image is pasted back everywhere outside the mask, with a soft edge a few pixels wide, so no seam shows. The result has the source image's size.
• The size and shape follow the photo. The edit keeps the photo's shape, and the node's Aspect Ratio is ignored. The edit renders at the size (1K, 2K or 4K) nearest the photo's, whatever size is set. For example, a 768 × 1024 photo runs at 1K, and a 3000 × 4000 photo runs at 4K. The edit is priced at that size. When the photo's size cannot be read, the size you set is kept.
• A mask costs nothing extra. The edit costs the same as the same edit without a mask at that size. The mask does not count as an input image.
• Grey is a partial change. A soft brush edge stays soft.
• PNG, JPEG or WebP only. A mask with Format set to GIF or SVG is refused before anything is charged.

Change the aspect ratio

To extend an image to a new shape, choose Nano Banana Edit, set the target Aspect Ratio, and use an instruction such as:

Start from a preset

Modify Image shares a set of transform presets with Generate Image. Open the preset menu on the node and choose one:

• Stylized Subject — change the style of the person but not the world around them, or the other way round. For example, a cartoon person in a real photo, or a real person in a cartoon world.
• Edits — remove or replace the background, colorize, restore an old photo, relight or restyle. Two Doodle Overlay presets draw marker doodles over the photo and leave the pixels underneath unchanged. Four photo-finish edits make a photo look less retouched: Add Skin Realism, Natural Photo Finish, Phone Camera Look and Undo Beauty Filter.

These presets keep their instruction in the pre and post text, so the prompt field stays free for your own words. See Presets.

Credits

The cost depends on the model and on its quality, resolution or rendering-speed setting. The Run button shows the exact price before you run. Flux 2 Max (Safety Tolerance) is priced by output size and by the number of reference images. MiniMax H3 Image and Seedream 5.5 are priced by size and by each input image, the source included. On GPT Image 2.6, a mask costs nothing extra. On GPT Image 2.6, PNG, JPEG and WebP cost the same. A GIF is priced by its quality and its band of frames, and an SVG costs the same image in PNG plus a fixed amount for its level of detail. See Formats in Generate Image. Each model page lists its prices.

Tips

• Describe the change, not the scene. "Make the sky purple" works better than "a landscape with a purple sky".
• Start with a low Strength. On models that have it, begin around 0.3 to 0.5 to keep the original structure, then raise it for bigger changes.
• Keep brand assets as files. A model redraws a logo slightly wrong. Edit the picture first, then place the real logo with Image Overlay.
• Chain simple steps. For example, remove the background, then upscale the cut-out, instead of asking one model for both.