What is ai-shortvideo-pipeline?

myAiVideos is an open-source project that automates the production of short-form video content in Chinese, from topic discovery to a publish-ready video. It works through seven stages: picking and ranking a trending topic, writing a script and visual prompts, generating images and video clips, producing audio narration, syncing everything in post-production, packaging for distribution, and scoring quality for future improvement. The backbone is a FastAPI service that orchestrates all seven stages, with an ARQ-based task queue for background work. A separate Java gateway built on Spring Boot handles authentication, routing between AI providers, circuit breaking via Resilience4j, and token and cost metering. If one AI provider fails, the gateway automatically rotates to the next one in the list. The project connects several AI models: DeepSeek and Zhipu GLM for text generation, Kling for turning prompts into images and video clips, Volcengine or MiniMax for speech synthesis, and faster-whisper for transcription. A CLIP model is used mid-pipeline to check whether generated images actually match their text prompts, and frames that fail this check are rejected before the video is assembled. Audio-video sync problems are handled by a four-tier rescue strategy that can adjust audio tempo, pad video segments, or rewrite narration rather than failing the job outright. A trace ID follows each request across the Java and Python layers, with Langfuse providing a full call-tree view for debugging. The stack is deployed with Docker Compose and includes a Vue 3 frontend for monitoring pipeline progress in real time via server-sent events. Credentials for each AI service are set in a .env file before starting. The project is released under the MIT license.