Seedream 5.0 Pro Interactive Editing: Point, Box, Replace, and Preserve

From “change something here” to a precise instruction

Seedream 5.0 Pro supports coordinate-aware editing. A user can click near an object or draw a box around a region, then combine that location with a natural-language instruction. This is useful for object replacement, local repainting, targeted removal, and edits across multiple reference images.

The two coordinate forms

• <point>x y</point> identifies a location and lets the model infer the surrounding object.
• <bbox>x1 y1 x2 y2</bbox> identifies a rectangular region from its top-left to bottom-right corner.

The values are normalized to a 0–999 coordinate space, so they are independent of the displayed pixel size. For a click at pixel position (x, y), divide by the displayed image width or height, multiply by 1000, and round to an integer.

normalizedX = round(clickX / displayedWidth * 1000)
normalizedY = round(clickY / displayedHeight * 1000)

Write the target and the operation together

Replace the object near image 1 <point>420 610</point> with a small ceramic lamp. Preserve the table, shadows, perspective, and wall color.

In image 1 <bbox>120 180 640 760</bbox>, replace the person on the left with the robot from image 2. Keep the person on the right unchanged.

If a box contains several subjects, name the intended one by position, clothing, color, or depth. Coordinates narrow the area; language resolves ambiguity.

Preservation instructions matter

A local edit can still affect nearby pixels if the prompt does not state what should remain fixed. Protect composition, faces, product geometry, lighting, reflections, typography, shadows, and camera perspective when they are important. For brand and e-commerce work, preservation rules are often more valuable than extra style adjectives.

Cross-image editing

For replacement across images, assign stable image numbers and specify which attributes transfer. “Use the red chair from image 2” is clearer than “make it like the other image.” If only shape and material should transfer, say that color, scale, or background should not.

Implementation checklist for a product UI

1. Record the rendered image position, displayed width and height, original dimensions, and image number.
2. Convert pointer coordinates from the screen or canvas into coordinates relative to the image.
3. Normalize each value to 0–999 and clamp it to the valid range.
4. Insert the point or box token beside the corresponding natural-language instruction.
5. Send only the images referenced by the final prompt and remap their image numbers consistently.

Further reading

Official Seedream 5.0 Pro interactive editing guide