Back to Details

rn-niulai-style-image

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

牛来风格图片生成

牛来风格图片生成

Use the runtime's image-edit tool for an existing still and its image-generation tool only for a new scene. Treat the task as 
style-transfer
when an edit target exists and as 
stylized-concept
for a new scene.
对于现有剧照，使用运行环境的 
image-edit
工具；只有在创作全新场景时，才使用 
image-generation
工具。存在编辑目标时，将任务视为 
style-transfer
；对于新场景，则视为 
stylized-concept
。

Choose the mode

选择模式

  • Default to 
    niulai-film-3d
    : rough low-budget 3D observed in public film frames.
  • Use 
    niulai-poster-ink
    only when the user explicitly asks for the water-ink poster. Never mix the poster with the film-frame references.
  • Describe the result as matching the observed public-frame visual grammar, not as an official art specification.
Read style-spec.md before authoring the prompt. Read evaluation-rubric.md before judging any output.
  • 默认使用 
    niulai-film-3d
    ：采用公开电影画面中可观察到的粗粝、低成本 3D 风格。
  • 仅当用户明确要求水墨海报时，才使用 
    niulai-poster-ink
    。绝不将海报模式与正片画面参考素材混用。
  • 将结果描述为匹配公开正片画面中可观察到的视觉语法，而不是官方美术规范。
编写提示词前，阅读 style-spec.md。评估任何输出前，阅读 evaluation-rubric.md。

Prepare inputs

准备输入

  1. Label the edit target as Image 1.
  2. Select two style references from 
    assets/style-reference/
    that cover face and body. The image tool accepts at most three inputs total.
  3. If a source still is wider than about 1600 px, resize a working copy first. Oversized stills fail.
  4. Preserve the target's character count, positions, action, clothing-color roles, key props, and camera crop unless the user asks to change them.
  5. Remove source logos, titles, dates, subtitles, and watermarks from the output.
  6. If a still is blocked by moderation, stop that still and switch to another adult scene. Do not paraphrase to retry. Prefer stills of adults; skip group shots of children.
  1. 将编辑目标标记为 Image 1。
  2. 从 
    assets/style-reference/
    中选择两张分别覆盖面部和身体风格的参考图。图片工具最多总共接受三个输入。
  3. 如果源剧照的宽度超过约 1600 px，请先调整工作副本的尺寸。尺寸过大的剧照会导致处理失败。
  4. 除非用户要求更改，否则保留目标中的角色数量、位置、动作、服装颜色及其角色区分、关键道具和镜头裁切。
  5. 从输出中移除源素材的 logo、标题、日期、字幕和水印。
  6. 如果某张剧照被审核机制拦截，停止处理该剧照，并换用另一个成人场景。不要通过改写提示词重试。优先选择成年人剧照；跳过儿童群体镜头。

Build the first prompt

构建首个提示词

Use this structure:
text
Use case: style-transfer
Asset type: <intended use and aspect>
Input images: Image 1 is the edit target; Images 2-N are style references. Match the look of Images 2-N, not a polished 3D reinterpretation.
Primary request: Rebuild Image 1 as cheap inflatable-costume 3D from an early-2000s amateur animation, while preserving <scene invariants>.
Character translation: Replace people with distinct upright bovine characters whose bodies are sausage-like inflatable suits; preserve identity through role, clothing color, pose, and left-right position rather than human facial likeness.
Style/medium: balloon torsos with almost no waist; oversized mask heads; pale plastic muzzles; small half-lidded eyes; uneven blunt horns; short tube limbs; mitten hands; white sock-hooves; blurry low-resolution fur-color bitmaps stretched across smooth blobby meshes; broad smears, visible seams and scale mismatch; weak contact shadows; primitive flat lighting; stock tutorial trees and painted backdrops; mild cinema-screen capture softness.
Constraints: <invariants>; no text, logo, watermark, subtitle, or date.
Avoid: professional 3D, clean low-poly, cinematic game art, realistic fur, appealing mascot design, Pixar/Disney, global illumination, correct human anatomy under an animal head.
Do not ask merely for "low poly." That usually produces clean modern game art and misses the target. The dominant miss is a well-made animal-headed human. If the body has a waist, articulated fingers, groomed fur, or cinematic light, the result is wrong.
使用以下结构：
text
Use case: style-transfer
Asset type: <intended use and aspect>
Input images: Image 1 is the edit target; Images 2-N are style references. Match the look of Images 2-N, not a polished 3D reinterpretation.
Primary request: Rebuild Image 1 as cheap inflatable-costume 3D from an early-2000s amateur animation, while preserving <scene invariants>.
Character translation: Replace people with distinct upright bovine characters whose bodies are sausage-like inflatable suits; preserve identity through role, clothing color, pose, and left-right position rather than human facial likeness.
Style/medium: balloon torsos with almost no waist; oversized mask heads; pale plastic muzzles; small half-lidded eyes; uneven blunt horns; short tube limbs; mitten hands; white sock-hooves; blurry low-resolution fur-color bitmaps stretched across smooth blobby meshes; broad smears, visible seams and scale mismatch; weak contact shadows; primitive flat lighting; stock tutorial trees and painted backdrops; mild cinema-screen capture softness.
Constraints: <invariants>; no text, logo, watermark, subtitle, or date.
Avoid: professional 3D, clean low-poly, cinematic game art, realistic fur, appealing mascot design, Pixar/Disney, global illumination, correct human anatomy under an animal head.
不要仅仅要求“low poly”。这通常会生成干净的现代游戏美术风格，并偏离目标。最常见的偏差是生成制作精良的兽头人身形象。如果角色身体有腰部、可活动的手指、经过修饰的毛发或电影级光照，结果就是错误的。

Iterate deliberately

有意识地迭代

  1. Generate one first-pass image.
  2. Compare target, style reference, and output with 
    scripts/.venv/bin/python scripts/make_comparison.py
    when local paths are available. Create 
    scripts/.venv
    and install Pillow if that interpreter is missing.
  3. Score all six rubric dimensions. Continue if total score is below 27/30, any dimension is below 4, or an automatic-fail condition applies.
  4. Change one dominant failure per iteration:
    • live-action leftover -> rebuild the whole frame as 3D, including walls, props, and fire;
    • too polished -> degrade texture scale, seams, lighting coherence, and capture sharpness;
    • too cute -> flatten expressions, shrink eyes, widen muzzle, stiffen pose;
    • mascot-in-a-photo or vinyl suit -> replace with balloon geometry and blurry stretched low-resolution color bitmaps;
    • ragdoll overshoot -> remove button eyes, Frankenstein stitches, and letter prints;
    • reference bleed -> do not use a leopard/print ref unless the target character is spotted;
    • generic low-poly -> replace crisp facets with blobby forms and mismatched pasted textures;
    • material still looks skilled -> remove embossed/procedural detail, flatten to blurry low-resolution bitmaps, broad UV smears, and simple smooth meshes;
    • scene drift -> repeat character count, positions, props, and action invariants;
    • character merging -> name every character by position and action.
  5. Re-state all invariants on every edit. Stop only after the threshold passes.
Keep iteration history under 
assets/showcases/iterations/
when building or validating this Skill. Save selected finals under 
assets/showcases/final/
.
  1. 生成第一版图片。
  2. 在本地路径可用时，使用 
    scripts/.venv/bin/python scripts/make_comparison.py
    对比目标图、风格参考图和输出结果。如果缺少该解释器，请创建 
    scripts/.venv
    并安装 Pillow。
  3. 对全部六个评分维度进行评分。如果总分低于 27/30、任一维度低于 4，或触发自动不通过条件，则继续迭代。
  4. 每次迭代只修改一个最突出的失败项：
    • 存在真人画风残留 → 将整个画面重建为 3D，包括墙面、道具和火焰；
    • 效果过于精致 → 降低纹理分辨率与画面采集清晰度，强化接缝并削弱光照一致性；
    • 造型过于可爱 → 让表情变得平板，缩小眼睛，拓宽口鼻，使姿态更加僵硬；
    • 像照片中的吉祥物或乳胶服装 → 使用气球状几何体，以及模糊、拉伸的低分辨率色彩位图；
    • 布娃娃感过于夸张 → 移除纽扣眼、弗兰肯斯坦式缝线和字母印花；
    • 参考素材串扰 → 除非目标角色带有豹纹或斑点图案，否则不要使用豹纹/印花参考图；
    • 呈现为泛化的 low-poly 风格 → 将锐利切面替换为团块状形体，以及不匹配的粘贴纹理；
    • 材质看起来仍然过于精致 → 去除浮雕/程序化细节，将其处理为模糊的低分辨率位图、大面积 UV 涂抹痕迹和简单平滑网格；
    • 场景发生漂移 → 重新强调角色数量、位置、道具和动作等不变量；
    • 角色相互融合 → 按位置和动作逐一命名每个角色。
  5. 每次编辑都要重新说明所有不变量。只有通过阈值检查后才能停止。
构建或验证此 Skill 时，将迭代历史保存在 
assets/showcases/iterations/
下。将选定的最终成果保存在 
assets/showcases/final/
下。

Deliver

交付

Report:
  • final image paths;
  • final prompt or prompt template;
  • per-image rubric score;
  • which iteration changed the result materially;
  • built-in image generation as the execution mode.
报告：
  • 最终图片路径；
  • 最终提示词或提示词模板；
  • 每张图片的评分量表得分；
  • 哪一次迭代带来了实质性改进；
  • 以内置图片生成功能作为执行模式。