O34 · Design Image Generation Serving (DALL-E) O34 · 设计图像生成服务（DALL-E）
Verified source经核实出处
OpenAI DALL-E public product. Asked at onsites per 一亩三分地 2024. Credibility B.
Architecture架构
flowchart LR Req --> GW GW --> MOD[Text Moderation] MOD --> SCHED[GPU Scheduler] SCHED --> DIFF[Diffusion fleet] DIFF --> UP[Upscaler] UP --> NSFW[Image NSFW Filter] NSFW --> BLOB[(Blob / CDN)] GW --> URL[(Return signed URL)]
Key decisions关键决策
- **Async with polling or SSE**: 10-30 s generations too long for synchronous.**异步 + 轮询/SSE**：10-30 s 对同步太长。
- **Batch multiple prompts per GPU step** to amortise diffusion overhead.**同 GPU 批量多 prompt** 摊销 diffusion 固定开销。
- **Two-stage filter**: moderate prompt before generation + classifier on output image.**两阶段过滤**：生成前审核 prompt + 对输出图做分类器检查。
- **Signed URLs via CDN**: users never hit GPU tier for downloads; TTL limits leakage.**CDN 签名 URL**：下载不打回 GPU；TTL 限泄漏。
Follow-ups追问
- Latency vs quality? expose steps/guidance params with documented cost tiers.延迟 vs 质量？暴露 steps/guidance 参数并标成本档位。
- Abuse prevention? per-account rate limits, fingerprint dedup against banlist.防滥用？账号级限流，指纹与黑名单比对。