A Seven-Layer Factory for AI Short Videos

What it does

myAiVideos is an open-source pipeline that automates the full short-video supply chain—from topic discovery and script generation to visuals, audio synthesis, post-production, and distribution—with FastAPI orchestrating the work and a Spring Boot gateway handling platform governance. It targets Chinese-language content formats such as news commentary and knowledge explainers, coordinating external models including DeepSeek, Kling, Zhipu GLM-4V, and Volcengine TTS to produce a publish-ready clip. The architecture is deliberately decoupled into seven independent layers so individual stages can be swapped without rewiring the entire flow.

The interesting bit

The project applies enterprise infrastructure patterns to generative media. A Java gateway uses Resilience4j to failover between LLM providers automatically, while the Python side enforces quality through CLIP-based image consistency checks, prompt anchoring to preserve visual identity across segments, and a four-tier strategy to rescue audio-video sync drift. Cross-language tracing via Langfuse and tenant-level metering round out the production-minded engineering.

Key highlights

• Seven decoupled pipeline layers (L1–L7) covering trending, creative, visual, audio, post-production, distribution, and optimization
• Multi-model failover with Resilience4j circuit breakers rotating across DeepSeek, Qwen, and GLM providers
• AI quality gating: prompt anchoring for visual continuity, CLIP text-image consistency rejection, and AV sync auto-rescue
• Full-stack observability with trace_id propagated across Java and Python services
• Metering and rate limiting implemented via AOP aspects, aggregated per tenant without code changes

Verdict

Development teams building automated content platforms will find a complete, opinionated reference architecture; casual users looking for a single-script video generator should look elsewhere.