Files
slide/decks/gen2D/pages/pipeline.md
T
wonder 19075592e5
Deploy Slides / build-and-deploy (push) Successful in 1m16s
fix: trim gen2D slide content to prevent overflow beyond header/footer
- Reduce Item count per slide: pipeline 4→3, solution 4→2, highlights 6→4
- Remove tech-stack architecture layering Item (moved to tech-stack grid)
- Simplify eino-graph: shorter Go code, 3 Items→2
- Simplify prompt-agent: smaller mermaid, shorter Item text
- Simplify async-task: merge 3 Items→2, remove footer
- Simplify data-model: merge 4 grid items→2
- Simplify architecture: reduce ASCII diagram from 38 to 20 lines
- Simplify dedup-ws: remove JSON code block, shorter Items
- Simplify deployment: smaller mermaid, shorter Items
- Simplify problem: shorter Item text, remove bullet lists
- Simplify style-system: fix typo, shorter Item text
2026-05-30 14:43:23 +08:00

28 lines
930 B
Markdown

# 核心生成管线
基于 Eino `compose.Graph` 编排的四阶段智能体协作管线
<Item title="① PromptOptimizer — 提示词优化">
合并工程风格 + 任务风格覆盖,生成三段式规范提示词。调用 Eino Chain 驱动 LLM 优化,无 Key 自动回退模板。
</Item>
<Item title="② AssetGenerator — 素材生成">
调用 OpenAI 兼容的文生图 API(默认 Stability AI),根据优化后提示词生成素材。
</Item>
<Item title="③ QualitySupervisor → ④ FormatAdapter">
视觉模型自动质检,不合格触发重优化(最多 3 次)。通过后组装 Spritesheet + 元数据,上传七牛云 Kodo。
</Item>
<div class="mt-4 text-sm text-gray-400">
```mermaid {scale: 0.7}
graph LR
A[用户输入] --> B[PromptOptimizer] --> C[AssetGenerator] --> D{质检}
D -->|通过| E[FormatAdapter] --> F[输出]
D -->|不通过| B
D -->|重试≥3| E
```
</div>