material-enhancement — 优化服装材质和细节质感

把一张面料糊掉的商拍图修成面料清晰真实的图，构图和人一动不动。

这是一个后处理技能：AI 生成的服装图最容易崩的就是面料——远看还行，放大一看针织变成一片糊。给它一张真实的高清商品图当参照，把表面重建回来。

生成效果示例

实际执行的命令：

dlazy gpt-image-2 \
  --prompt 'Texture enhancement pass. Image 1 is the on-model photo to improve. Image 2 is the high-resolution product photo that defines the true fabric. Rebuild the sweater surface in image 1 using the real texture from image 2: crisp cable-braid relief, visible knit loops and yarn twist, natural wool loft, correct fold shadows and a matte fibre sheen. Do not change anything else — the model face, hair, hands, pose, brown trousers, background, framing and colour grading must stay identical to image 1. The garment silhouette and colour must not shift; only the material fidelity and micro-detail improve. Photorealistic, sharp, no text, no watermark.' \
  --images docs/material-enhancement/source-image.jpg docs/material-enhancement/hires-product.jpg \
  --size 1024x1536 --quality high --imageFormat jpeg \
  --save docs/material-enhancement/example-output.jpg

输出

example-output.jpg — 1024×1536，60 credits。麻花辫的立体起伏、菱形提花的凹凸、羊毛的绒毛感与褶皱处的暗部层次被重建；模特五官、发型、手部、姿势、棕色长裤、灰墙背景与整体色调保持不变，毛衣轮廓与军绿色未偏移。

1、能力边界

不做：不改服装轮廓与颜色；不改模特与背景；不把一种面料换成另一种（换面料请用 fabric-on-body）。

2、输入素材规则

生成前先自检这几条硬性约束：

• 大小：20KB ~ 15MB
• 分辨率：大于 400×400
• 格式：jpg / jpeg / png / webp

原图（✅）：构图与人物都满意、只有面料不行的图。

高清商品图（✅）：同一件商品的高分辨率平铺图或特写，能看清织法。

注意事项

3、材质增强的两条铁律

铁律一：只改表面，别的一个像素都别动。

Do not change anything else — the model face, hair, hands, pose, other garments,
background, framing and colour grading must stay identical to image 1.
The garment silhouette and colour must not shift; only material fidelity
and micro-detail improve.

铁律二：把「好质感」翻译成具体的表面特征，否则模型只会整体加锐化。

再加一条褶皱层次：correct fold shadows with proper ambient occlusion in the creases——这是「看起来有厚度」的来源。

4、工具调用

本技能使用 dLazy 的 gpt-image-2（图像编辑模型 + --quality high；材质增强要求「几何与色彩零变化、表面高频信息重建」，是典型的双图参考定向修复）。

调用方式

两种等价写法，选一种。统一入口会自动选后端、失败重试、建目录落盘、估算成本：

# A. 统一入口（推荐）：可切任意后端，加 --dry-run 不计费空跑
node scripts/gen.mjs --task material-enhancement \
  --prompt '<见下方 Prompt 模板>' \
  --images <按下表顺序> \
  --save output/material-enhancement-<sku>.jpg

# B. 直接用 dLazy CLI（不想引入 Node 依赖时，效果等价）
dlazy gpt-image-2 --prompt '...' --images ... --save output/material-enhancement.jpg

参数约定（本技能固定用法）

Command Examples

# basic call
dlazy gpt-image-2 \
  --prompt 'Texture enhancement pass. Image 1 is the on-model photo to improve. Image 2 is the high-resolution product photo defining the true fabric. Rebuild the garment surface in image 1 using the real texture from image 2: crisp knit relief, visible loops and yarn twist, natural loft, correct fold shadows. Do not change anything else — model, pose, background, framing and colour grading stay identical to image 1. Only material fidelity improves.' \
  --images docs/material-enhancement/source-image.jpg docs/material-enhancement/hires-product.jpg \
  --size 1024x1536 --quality high

# complex call: 批量给一组生成图做材质后处理
HIRES=docs/material-enhancement/hires-product.jpg
LOCK='Do not change anything else — the model face, hair, hands, pose, other garments, background, framing and colour grading must stay identical to image 1. The garment silhouette and colour must not shift; only material fidelity and micro-detail improve. Photorealistic, sharp, no text, no watermark.'
FEAT='crisp cable-braid relief, visible knit loops and yarn twist, natural wool loft, matte fibre sheen, correct fold shadows with proper ambient occlusion in the creases'
for f in docs/material-enhancement/raw/*.jpg; do
  dlazy gpt-image-2 \
    --prompt "Texture enhancement pass. Image 1 is the on-model photo to improve. Image 2 is the high-resolution product photo that defines the true fabric. Rebuild the garment surface in image 1 using the real texture from image 2: $FEAT. $LOCK" \
    --images "$f" "$HIRES" \
    --size 1024x1536 --quality high --imageFormat jpeg \
    --save "docs/material-enhancement/enhanced/$(basename $f)"
done

# 先估价不真跑
dlazy gpt-image-2 --dry-run --prompt '...' --images a.jpg b.jpg --size 1024x1536 --quality high

延伸阅读

5、Prompt 模板

Texture enhancement pass.
Image 1 is the on-model photo to improve.
Image 2 is the high-resolution product photo that defines the true fabric.

Rebuild the [服装部位] surface in image 1 using the real texture from image 2:
[第三节表格里对应面料的具体特征],
correct fold shadows with proper ambient occlusion in the creases.

Do not change anything else — the model face, hair, hands, pose, other garments,
background, framing and colour grading must stay identical to image 1.
The garment silhouette and colour must not shift; only material fidelity
and micro-detail improve.

Photorealistic, sharp, no text, no watermark.

按问题追加的修正句

6、执行流程

1. 确认两张图是同一件商品——这是最容易犯的错。
2. 判断原图值不值得救：结构崩坏的重跑生成，只有面料糊的才用本技能。
3. 写面料特征：查第三节表格，用具体特征替代「质感更好」。
4. 写两条铁律：只改表面 + 别的不许动。
5. --quality high（必须） → --batch 2 挑图，落盘到 docs/material-enhancement/。
6. 并排对比原图与输出：面料是否真的重建了（不是加锐化）、脸和背景是否没动、颜色是否没偏。

7、常见问题

Tips

Visit https://dlazy.com for more information.