Edit an Image with Magic Edit

Magic Edit changes an existing image from written instructions. Use it for requests such as changing the background, adding an object, or trying a new look while keeping a subject. If you only need a new image from text, start with Your First Image.

Follow the editing flow

  1. Open Magic Edit and choose a Base Model and Checkpoint in Model. The available models are Qwen Image, Qwen Image 2.1, and Flux2 Klein.
  2. In Input Images, upload or select the image you want to edit. You can crop it before submitting. The first image determines the output ratio and affects the price, so check its crop and order.
  3. Write an Edit Prompt that says what to change and what to keep.
  4. Review Settings, especially Private Mode and Hide Prompt, then select Edit Image.
  5. Follow the job in Results. On mobile, open the Results tab. Open the finished image to download it or use Recreate for another attempt.

Start with one image and a specific request:

Change the overcast sky to a warm sunset. Keep the red fox, its pose, the stream, and the camera angle unchanged. Match the new light on the water and fur.

If too much changes, shorten the request and name the details that must stay. If the edit is too subtle, describe the intended new appearance more concretely. Use Use Fixed Seed when comparing two versions of the prompt.

Edit with multiple images

The supported number of inputs depends on the model: Qwen Image accepts up to three selections, Qwen Image 2.1 up to five, and Flux2 Klein up to three. Drag the images to set their order. Their order matters to the prompt.

  • Qwen Image: the selected references are combined left to right into one guide image. Refer to their visible positions, such as “the person on the left” and “the jacket on the right.”
  • Qwen Image 2.1: refer to distinct inputs with <image1> through <image5> as shown by the editor.
  • Flux2 Klein: describe the role of each reference in order, such as “use Reference 1 for the subject and Reference 2 for the clothing.”

Example with two references:

Keep the subject and pose from Reference 1. Use the blue jacket from Reference 2, matching its fabric and collar. Keep the original background and lighting.

The available model can change how closely details are preserved. Review the output instead of assuming that an instruction will be followed exactly.

Add a LoRA only when needed

Use Add LoRA to select a LoRA compatible with the chosen model. LoRAs can influence a style or concept; their weights can be adjusted or an entry disabled. If an edit becomes distorted or ignores the original image, retry without the LoRA or lower its weight. The editor supports up to five selected LoRAs.

For a head swap using the BFS LoRA, follow the separate Face Swap Guide. For normal face replacement, the Face Swap tool is a simpler starting point.

What affects cost?

The first reference sets the output width and height by default. Qwen Image 2.1 rounds each side to the nearest 32 pixels. Enable Scale to 2K in Settings to target a size close to the model's 2K image-generation sizes while keeping the first image's ratio. If you cropped it, the cropped ratio is used. The first image is scaled for the edit without padding or trimming; other references are not scaled to 2K. Images with extreme aspect ratios may have fewer pixels than the usual 2K output because the longer side is capped. An image already at a supported 2K size stays at that size. The output size may increase or decrease, so the credit cost may change. Flux2 Klein and Qwen Image 2.1 can also add a charge for each reference. See the current pricing page before submitting.