Function-style prompt + 数字权重 layer + 跨域 translation + meta-prompting → 2x2 grid 4 个 creature variants 带 per-cell label · 一次出 4 个 viral concept
项目内首验证 2026-05-01 · 单样本 · 7 unique 维度全 PASSED · 出片 Michelin-quality. 首样本: 4 extinct animals (Triceratops/T-Rex/Mammoth/Pteranodon) as sushi sculptures · 2x2 grid · 全标签. **3 个 framework-级新发现:** (a) **gpt-image-2 honor 显式 function-syntax prompt** · `Function NameXXX($Arg1, $Arg2)` + numbered steps + 权重 ::N + Negative ::-1 这套 LLM 风格 · 在 image gen 上有效 · 不是被忽略当 keyword bag · 启示: 复杂 multi-stage creative task 可以用 function 包装 vs flat description (b) **数字权重 ::N 作 priority hint 有效** · ::5 (Anchor) · ::4 (Morphology) · ::3 (Material) · ::2 (Light) · ::1 (Render) · ::-1 (Negative) · 不像 SDXL 那么强权重 · 但**层级感清晰** · 高权重 anchor 更显眼 · 启示: 跟 显式 'negative prompt:' (cinematic motion 验) 同属 prompt-engineering primitive 拓展 (c) **2x2 + 4 short labels (≤20 chars) ALL PASS** · 之前 Noor 'Noor 🌸' (5 chars) 经验上限上调 · 现在: "TYRANNOSAURUS REX" (17 chars) 全 readable + 拼写正确 · 但**仍是 grid 内多 short label · 不是 long-form text · 不要 extrapolate 到段落级** 其他 PASSED 维度: · Meta-prompting: model 自完成"分析 → 推断"两步 (auto-enumerate 4 iconic + auto-map 食材) · Cross-domain translation: animal anatomy ↔ sushi ingredients (mammoth fur → rice grains 完美) · 2x2 grid: cells 均匀 · production sweet spot (47s vs 5x5 14min) · Auto-enumeration: model 偏向 high-recognition subjects (跟 5x5 fox 同模式) Production use: · 餐饮品牌 marketing · 创意菜单 mock · viral SNS 4-cell 系列 · pre-viz 4 concept 看哪个最 viral · template 化: 替换 "extinct animals" + "sushi" 为其他 cross-domain pair (e.g., 神话动物 × 乐高 / 童话角色 × bento) Sister storyboard-5x5-style-grid (同 grid axis · 5x5 是 style 维 · 此条是 subject 维 4 variation). Sister templates-ink-wash-painting (同 placeholder template · 但 ink wash 单 subject · 此条 grid). Sister templates-surreal-mural-portrait (同 in-image text · 但 mural 是单 string · 此条 4 labels).
date: '2026-05-01T13:29:20+08:00'
result: pass
prompt:
text: >-
<instructions>
Input = 4 iconic extinct animals
Act as a Master Sushi Chef.
1. Analyze the Creature to find its Body Parts (e.g., Wings, Scales, Claws).
2. Infer the Sushi Ingredients to mimic those parts (e.g., Scales = Cucumber/Avocado; Wings =
Salmon Sashimi).
Function Sushi_Monster ($Creature, $Ingredients)
Anchor: [Edible Sculpture of $Creature] :: [Fresh Sashimi and Rice]::5
Morphology: Culinary creature design, the anatomy of [$Creature] constructed entirely from
[$Ingredients], rice grains forming the muscle mass, seaweed (Nori) forming the outline/details,
sashimi slices forming the armor/feathers::4
Material Physics: Glistening raw fish texture, sticky rice, wet surface tension, appetizing food
colors, organic soft-body physics::3
Illumination: Bright sushi bar lighting, clean wood reflections, appetizing gloss::2
Render Stack: Macro food photography, 8k, sharp focus, Michelin plating aesthetic::1
Negative: [Real animal, scales, fur, feathers, dry, inedible, plastic, messy]:: -1
Output: 2x2 grid. Execute Function for 4 variations of the creature. Name of the Creature
underneath.
</instructions>
refs: []
provider:
id: gpt_image_2
relay: apimart
config:
aspect_ratio: '1:1'
size: '1:1'
'n': 1
output:
path: ./sushi_monster_extinct_v1.png
bytes: 2305723
wall_seconds: 46.7
task_id: task_01KQH056E732CBV7HB527XTREQ
script: experiments/sushi_monster_test/test_v1_extinct.py
cost_yuan: 0.5
notes: >
Function-syntax prompt · weighted layers (::5..::-1) · 2x2 grid · 4 extinct animals as sushi.
46.7s wall · 出片 Michelin-quality 食物摄影级.
7 unique 维度全 PASSED:
(1) **Function-syntax prompt** · HONORED
· `Function Sushi_Monster($Creature, $Ingredients)` 被当作 instructional 框架
· Model 不只读 keywords · 是真的执行了"分析 → 推断 → 生成"步骤
· 形似 LLM agent 的 prompt structure 在 image gen 上有效
→ gpt-image-2 能 honor 显式 function/instruction 风格 prompt
(2) **Weighted layers (::5/::4/::3/::2/::1/::-1)** · STRONG · 可见 hierarchy
· Anchor (5) Edible Sculpture · 最强 honor (Mammoth 最纯 sculpture · T-Rex 较 textural)
· Morphology (4) `rice=muscle / nori=outline / sashimi=armor` · 完美执行 (Mammoth 米粒清晰可数)
· Material Physics (3) glistening 鱼皮光泽 visible
· Illumination (2) 木台反光 + 暖灯
· Render Stack (1) Michelin plating · wasabi/garnish 全在
· Negative (-1) `real animal/fur/dry` · 0 leakage
→ 数字权重作 priority hint 在 gpt-image-2 上有效 · 即使不像 SDXL 强权重
(3) **Meta-prompting** · PERFECT · model 自完成两步推理
· Step 1 自选: Triceratops · T-Rex · Woolly Mammoth · Pteranodon (经典 4 种)
· Step 2 自映射:
- Triceratops: salmon body + nori horns + rice details
- T-Rex: tuna red scales + nori lines + rice teeth
- Mammoth: rice grain "fur" (每粒可见) + sashimi underside
- Pteranodon: salmon wings + nori frame
→ 模型可被赋予"内部决策"任务 · 不只 keyword pattern matching
(4) **Cross-domain semantic translation** · EXCELLENT
· Mammoth fur → individual rice grains 排列 (creative 解)
· Triceratops frill → salmon arc with nori spots
· T-Rex teeth → small rice clumps
· Pteranodon wing membrane → smooth salmon slice
→ 跨域 anatomy ↔ ingredient 翻译能力强 · 不是简单 texture overlay
(5) **2x2 grid** · PERFECT · 项目内最小 grid · cells 比例均匀
→ 4-cell 是 production sweet spot (vs 5x5 14min · 2x2 47s)
(6) **Per-cell labels** · ALL 4 PRESENT and READABLE
· "TRICERATOPS" / "TYRANNOSAURUS REX" / "WOOLLY MAMMOTH" / "PTERANODON"
· serif typeface · 一致 gold/cream 色 · 一致 position (cell 底部)
· 拼写全对 · 包括 "TYRANNOSAURUS REX" 17 chars 长 string
→ **NEW finding**: 2x2 grid + 4 short label (≤20 chars) ALL PASS · 之前 Noor 经验 ≤5 chars · 上限上调
(7) **Auto-enumeration** · STRONG
· "iconic extinct animals" → 选了 4 最 iconic (T-Rex/Triceratops/Mammoth/Pteranodon)
· 没选 obscure 的 (Quagga / Thylacine 等)
· model 默认偏向 high-recognition subjects · 跟 5x5 fox 同模式
Production use:
· 餐饮品牌 marketing (sushi 店 menu / 创意菜单)
· educational content (creature anatomy as food)
· viral SNS 系列 (4 cells 易于 social 分享)
· 替换 "extinct animals" 为其他 theme · template 化潜力高
· pre-viz: 用此格式快速 mock up 4 个 concept 看哪个最 viral
Sister storyboard-5x5-style-grid (同 grid axis · 但 5x5 是 style enumeration · 此条是 cross-domain
creature variants).
Sister templates-ink-wash-painting (同 template-with-placeholders + 显式 function · 但 ink wash 是
single subject · 此条是 4-variation grid).
recipes/image_gen/gpt_image_2/prompts/.