ai-shortVideo-pipeline

ai-shortVideo-pipeline is one of the most thoughtfully engineered open-source video-generation projects we've encountered. Rather than a single script that calls one model, it's a full platform: a FastAPI orchestration core handling jobs via ARQ workers, a Spring Boot gateway managing auth, routing, circuit breaking, and metering, plus a Vue.js frontend. The seven-layer architecture—topic, creative, visual, audio, post-production, distribution, optimization—is genuinely decoupled, so you can swap in a better TTS engine or a different video model without refactoring everything. The engineering depth shows in the details. Multi-model failover with Resilience4j is not just a retry loop: it actively detects a failing provider and routes around it, which matters when you're running high-volume production. CLIP text-image consistency scoring catches off-prompt keyframes before they waste compute downstream. The AV sync auto-rescue with its four-tier strategy (audio tempo, video pad, narration rewrite) is a real pain point solved—manually fixing audio-video drift is a huge time sink in DIY pipelines. Langfuse trace propagation across Java/Python makes debugging cross-language calls tractable, and the AOP-based metering captures costs per tenant without code changes. Where it fits: developers building Chinese-language content pipelines, teams that want to own their video infrastructure, AI engineers experimenting with multi-model orchestration. The fit is real because the model choices—DeepSeek, Qwen, GLM, Kling v2.5, Volcengine/MiniMax TTS—are optimized for the Chinese content ecosystem, which narrowly-targeted SaaS often ignores. Where it doesn't fit: if you're a creator who wants to paste a script and get a video without touching Docker, this will frustrate you. There's no hosted version, no one-click cloud deploy, and you must provision GPU/TPU resources and API keys yourself. The pipeline is batch-oriented, so it's not for real-time video needs. Distribution is a layer you wire up yourself—there's no built-in TikTok or YouTube publishing. If you're on a team without DevOps capacity, the infrastructure overhead (Postgres, Redis, MinIO, two runtimes) is substantial relative to what you get out of the box. Compared to closed SaaS like Pictory or InVideo, you trade setup convenience for control, cost predictability (your API bills only), and fault tolerance. If you can self-host, this is the most robust open-source option we've seen for high-volume short-video production. If you can't, look elsewhere.