AI E-commerce Video Generation System
Chapter 01
Overview
An AI generation system for e-commerce short videos. Target pipeline: analyze the shots, transitions and plot structure of a reference video, write a script modeled on it (optionally with product placement), plan the storyboard, generate character images and segmented videos (including dialogue), and finally assemble the finished video automatically and generate a cover.
Chapter 02
Sample
Sample produced: the first segment of the original video was analyzed and split into 2 units at the original cut points; using 2 new character images generated by the system and the corresponding clips of the original as references, the video model generated the visuals and dialogue, and the system automatically assembled them into a 12-second clip. The lines are taken from the original video, and no product is placed; the subtitles in the frames were generated by the video model in imitation of the reference clips and are not a system subtitle feature.
01Clip generated by the system (in-development sample)
Chapter 03 03 items
Key points
-
Reference video analysis: in addition to shot detection, adds local gradual-transition detection (FFmpeg decoding, no extra model required), segments by scene and plot and labels the relationships between segments, so that one plot beat is not broken up by shot cuts.
-
Pipeline design: script imitation, storyboard planning, character images, segmented video generation, automatic assembly of the finished video and cover; every step can be reviewed manually.
-
Engineering foundation: multi-queue asynchronous tasks, call budgets and accounting, object storage upload sessions, permissions and capacity reclamation; real model calls must be explicitly enabled to avoid incurring charges by mistake.
Chapter 04 12 items
Tech stack
- Python 3.12
- FastAPI
- SQLAlchemy
- Celery
- PostgreSQL
- Redis
- FFmpeg
- React
- TypeScript
- Ant Design
- Volcano Engine Ark Seedream / Seedance
- DeepSeek / GLM / MiniMax
