Sophnet Image Edit

SophNet Image Edit

Overview

Edit existing images with Python scripts that handle task polling and structured output.

Script responsibilities:

• edit_image.py: core API caller and polling loop, outputs INPUT_IMAGE_COUNT, TASK_ID, STATUS, and IMAGE_URL.
• edit_and_preview.sh: wrapper for local use, calls edit_image.py, downloads first result, adds PREVIEW_PATH.

When to Use

• User asks to edit or transform existing images with Sophnet.
• Task uses one or more source images (URL or uploaded local files).
• Task depends on image order/role, such as:
  • put part of image-2 into image-1
  • render image-2 with image-1 style
  • blend/composite many images into one result
• Do not use when the task is pure text-to-image generation; use sophnet-image-generate.

Image Path Resolution

When users upload images in chat channels, files are usually saved under media/inbound/images/ in the workspace. Use Media Understanding logs to get exact resolved paths.

Typical log pattern:

[Media Understanding] Resolved relative path: "media/inbound/images/xxx.jpg" -> "/absolute/path/to/workspace/media/inbound/images/xxx.jpg"

Path rules:

• --image accepts URL, data URI, or local file path.
• Local file paths are auto-converted to data URI by the script.
• For multi-image tasks, pass every image explicitly and in required order.

Multi-Image Rules

1. Keep source image order stable and explicit.
2. Bind prompt language to order (image-1, image-2, image-3...).
3. Do not drop "reference" images; pass all images used by the prompt.
4. Verify script output INPUT_IMAGE_COUNT equals expected image count before trusting results.

Recommended prompt pattern for multi-image tasks:

• 图1作为底图，图2提供主体元素，图3提供风格参考；将图2主体抠出放到图1右上角，并按图3风格整体渲染。

Quick Reference

Notes:

• --dry-run does not require API key and is for input-count verification only.
• Prefer --images-file when handling many images or values that may contain commas.

/tmp/image-list.txt format example:

# One image reference per line
media/inbound/images/base.jpg
media/inbound/images/object.png
https://example.com/style.jpg

Script Output Fields

The script outputs structured key=value lines. Parse all of them:

Presenting Results to Users

After a successful edit, present the results with context -- not just a bare URL. Include:

1. What was done: mention the edit instruction / style applied.
2. Input/output image counts.
3. The result image URL(s): provide the IMAGE_URL so the user can view/download.
4. If PREVIEW_PATH is present, mention the local file path.

Example response pattern:

图片编辑完成！

- 编辑指令：把图片变成水彩画风格
- 输入图片数：1
- 结果图片：<IMAGE_URL>

Implementation

1. Resolve all uploaded image paths from Media Understanding logs.
2. If this edit step follows sophnet-image-generate, prefer upstream IMAGE_URL for handoff; use PREVIEW_PATH only when local-file input is explicitly intended.
3. Build prompt with explicit image order semantics.
4. Run script with one --image per image, or --images, or --images-file.
5. For multi-image tasks, run once with --dry-run and confirm INPUT_IMAGE_COUNT matches intended image count.
6. Parse ALL output fields (see table above) and present them to the user.
7. Use PREVIEW_PATH when present.

API payload shape:

• model: Qwen-Image-Edit-2509
• input.prompt: edit instruction with image-role semantics
• input.images[]: all source image refs in strict order
• parameters: size, n, watermark

Common Mistakes

• Passing only one image for a multi-image prompt.
• Mentioning image-2/image-3 in prompt but not passing those images.
• Reordering --image arguments unintentionally.
• Ignoring INPUT_IMAGE_COUNT mismatch.
• Using non-existent local paths.
• In generate→edit workflows, defaulting to PREVIEW_PATH instead of IMAGE_URL without reason.
• Only showing a bare URL without context -- always present a summary with edit details.