🎯 场景：给两张前后对比图，反向生成"如何从 before 到 after"的编辑指令。用于训练 image-editing 模型的 SFT 数据建设、Photoshop 操作教学、design 版本对比说明。需要 vision 模型。

Quick Use

Use when: You have a before/after image pair and want to generate the natural-language edit instruction that would produce the change — for image-edit-model training data, design diffs, or tutorial generation. Fill in: {{before_image}} = original; {{after_image}} = edited; {{instruction_style}} = natural / technical / minimal. You'll get: The edit instruction in the requested style, list of detected changes, and complexity rating. Output is JSON. Requires a vision-language model.

Purpose

Reverse-engineer the edit instruction from a before/after image pair. Used in training data construction for image-editing models (e.g. InstructPix2Pix-style), design version comparison documentation, and tutorial generation. Output structured so simple vs complex edits can be sorted, and so the same pair can yield instructions in multiple styles.

Prompt

You analyze a before/after image pair and generate an edit instruction
that would produce the change.

Before image: {{before_image}}
After image:  {{after_image}}

Instruction style: {{instruction_style}}

Style meanings:
- "natural"   : How a user would describe the edit ("make the sky
                 sunset", "add a hat on the dog").
- "technical" : Photoshop / image-editor terminology ("apply
                 gaussian blur radius 5px", "increase saturation
                 +30").
- "minimal"   : Shortest unambiguous instruction.

Steps:
1. Identify what changed. Compare globally then per-region.
2. Categorize the changes:
   - "object_added"      : New object/element in after.
   - "object_removed"    : Object/element in before but not after.
   - "object_modified"   : Object present in both, attributes changed.
   - "color_or_lighting" : Color / saturation / brightness / shadow
                            changes, no object change.
   - "geometric"         : Crop / resize / rotate / perspective.
   - "style_or_filter"   : Artistic style change (e.g. photo →
                            painting).
   - "background_only"   : Background changed, foreground same.

3. Generate the instruction in the requested style.

4. Rate complexity: "simple" (1-2 changes, 1 type) / "compound"
   (multiple changes coordinated) / "complex" (many changes, hard
   to articulate fully).

Return ONLY this JSON object:
{
  "detected_changes": [
    {"category": "<from list>", "description": "<short>"}
  ],
  "instruction": "<the generated instruction in requested style>",
  "alternative_instruction_minimal": "<the same edit in minimal style for reference>",
  "complexity": "simple" | "compound" | "complex",
  "ambiguity_note": "<if change could be described multiple equally-valid ways>",
  "decision_basis": "<one sentence, <=30 words, no internal CoT>"
}

Example

Before: photo of a dog. After: same photo, dog is wearing a red bow tie.

Expected output:

{
  "detected_changes": [
    {"category": "object_added", "description": "Red bow tie added on the dog's neck."}
  ],
  "instruction": "Add a red bow tie on the dog.",
  "alternative_instruction_minimal": "Add a red bow tie.",
  "complexity": "simple",
  "ambiguity_note": "",
  "decision_basis": "Single object addition, clearly identifiable; instruction is unambiguous."
}

Failure Modes

• Phantom changes — model claims a change that isn't actually present. Validate by carefully comparing before/after; don't trust the model's claims uncritically.
• Missed subtle changes — changes in lighting / minor color shifts not detected. Sample edge cases.
• Style drift — natural style outputs Photoshop language or vice versa. Sample by style and verify match.
• Compound under-detection — multi-change edit collapsed to one description. Check detected_changes list size on known compound edits.
• Position / size hallucination — model says "moved the dog to the left" when dog stayed in place but background shifted. Spatial changes are hard; mark high complexity.

Tuning Notes

• 模型差异：strong VLM 必须的。中档 VLM 在精细 color / lighting changes 上失败率高。
• 温度：0.0–0.3。
• 与 multimodal/vlm-image-description-verifier 的关系：那张卡审 caption 是否匹配图；本卡审两图差异。语义不同。
• 与 multimodal/structured-caption-generator 的关系：那张卡描述单 图；本卡比较两图。
• 用作 image-edit 模型 SFT 数据：(before_image, instruction, after_image) 三元组就是 training pair，本卡产 instruction。建议 跑前用 image-similarity 先粗筛"有显著差异"的对，否则会产生大量 trivial instructions。
• 高分辨率 + 同源拍摄敏感：训练数据要求两图除编辑外完全一致；不同 曝光 / 不同角度 / 不同压缩 artifact 会让 instruction 含噪声。

Changelog

• 0.1.0 — Initial card.