单 subject + 25 cells × distinct styles · style-axis (vs narrative/rhythm-axis) · style 探索 / 客户教育 / pre-viz · minimal prompt unlock model prior
项目内首验证 2026-05-01 · 最大 grid (25 cells) + 最短 prompt (10 词). 新 axis: style-axis (vs narrative-axis 3x3 · rhythm-axis 12-panel · time-axis 16-panel daycycle). Subject 锁定 · style 变化 · 同 fox 25 种 anime/cartoon/illustration 风格. 5 unique 维度 (全 PASSED 4/5 · 1 NO): (1) 5x5 layout · 25 cells · model 上限 ≥25 · cells 比例均匀 (2) Minimal prompt → rich output · prompt-is-coordinate 经典 demo (3) ~18-20 RECOGNIZABLY distinct styles (5 partial overlap on cyberpunk/cute mascot) (4) Subject auto-consistency (model 选 fox + 锁 waving pose) (5) Auto-labels NO · 无文字标签 (要 label 须显式 'with style names below') **GOTCHA: 14 min wall** (855s actual vs 100s estimate · 8.5x) · 25-cell grid 是 ApiMart/OpenAI 资源密集 · queueing 严重 · cost still ¥0.5 但 wall 30-50x 普通 generation · ApiMart polling 8s + 360s deadline 对此 task 不够 · production 跑此 cap 须 timeout ≥20min Production use: · style 探索工具 · ¥0.5 看 model 25 种风格 prior on a subject · 客户教育 ('我们风格库' baseline) · 风格对比研究 · pre-viz 风格选项 Sister storyboard-3x3-narrative (axis: narrative · 单 subject 不变 · 9 个 narrative beat) + storyboard-12panel-rhythm (axis: rhythm · brand reel · 12 panel asymmetric).
date: '2026-05-01T08:14:00+08:00'
result: pass
prompt:
text: create a 5x5 grid of distinct animation styles
refs: []
provider:
id: gpt_image_2
relay: apimart
config:
aspect_ratio: '1:1'
size: '1:1'
'n': 1
output:
path: ./animation_styles_5x5_v1.png
bytes: 1971739
wall_seconds: 855
task_id: task_01KQGDBCVWWZX6N9DSCAKJJF5Z
script: experiments/animation_styles_grid_test/test_v1.py
cost_yuan: 0.5
notes: >
5x5 grid (25 cells) · MINIMAL prompt (10 词 / 46 chars) · 14 min wall (8.5x ApiMart 100s estimate ·
5x5 stress test).
5 unique 维度结果:
(1) **5x5 grid layout (25 cells)** · PERFECT
· 项目内最大 grid (vs 3x3=9 / 4x4=16 / 4x3=12 / 16-panel daycycle)
· cells 比例均匀 · 良好 separator
→ gpt-image-2 grid 上限 ≥25 cells
(2) **Minimal prompt** (10 词) · STRONG demo of prompt-is-coordinate
· 无 subject 描述 · 无 style 列表 · 无 layout 详细 → 仍出 rich grid
· model 完全靠 prior 填空 (auto-pick subject + auto-enumerate styles)
→ memory/feedback_prompt_is_coordinate_not_script.md 实证
(3) **Style enumeration** · ~18-20 RECOGNIZABLY distinct (out of 25)
ROW 1: Pixar 3D · rubber-hose (Mickey) · 90s Disney 2D · stop-motion felt · mid-century cut-paper
ROW 2: CN 90s flat · DreamWorks 3D · photoreal CG · botanical illustration · atomic mid-century
ROW 3: anime · film noir B&W · vintage rubber-hose B&W · psychedelic Lisa Frank · plush/Muppet mascot
ROW 4: flat vector · matte concept art · pencil line drawing · cyberpunk wireframe · storybook illustration
ROW 5: surrealism · graphic novel · kawaii chibi · sumi-e ink wash · cyberpunk neon
· ~5 cells partial overlap (2 cyberpunk · 2 cute mascot · 2 vintage B&W)
→ model 风格词汇丰富 · 但 25 cell 拉满会有重复
(4) **Subject consistency** · PERFECT
· 同 fox 角色 (waving pose) 贯穿 25 cells · 仅 style 变
· model 自选 simple anchor (fox = 跨风格易于呈现的 cute mascot)
→ minimal prompt → 模型默认锁单 subject 让 style 成比较轴
(5) **Auto-labels** · NO LABELS
· 干净 grid · 无文字标签
· 显式要 label 须在 prompt 加 'with style names below each cell'
Production use:
· style 探索工具 (¥0.5 看模型对一类 subject 的 25 种风格 prior)
· 客户教育素材 ('我们风格库长这样 · 选一个')
· 风格对比研究 ('同 subject 不同 style' baseline)
· pre-viz 风格选项 (vs 之前 3x3/4x4 narrative-axis · 此条 style-axis)
Cost note: ¥0.5 不变 · 但 wall 14min vs typical 40s · 暗示 ApiMart/OpenAI 对高 cell-count grid 排队/资源更紧.
Production 不要批量跑 5x5 · 单次探索可以.
recipes/image_gen/gpt_image_2/prompts/.