Alibaba's Pixelle-Video Turns Any Topic Into a Finished MP4 for Free
Pixelle-Video is a free, open-source pipeline that turns a single text topic into a fully edited short video — script, visuals, voice, and music included
ProTakeaways
- What it is: Pixelle-Video is a free, open-source engine that converts a text topic into a finished short video with script, visuals, voice, and music.
- How it works: An 8-step async pipeline chains an LLM, image/video generation (FLUX, WAN 2.1), TTS, HTML template rendering via Playwright, and ffmpeg composition.
- Fully modular: Every component — LLM, image model, TTS engine, video template — is swappable via ComfyUI workflow JSON files or direct API providers.
- Cost: Free with local GPU (Ollama + ComfyUI); ~$0.01/video with Qwen API; ~$0.50–$2 with full cloud stack (GPT-4o + RunningHub).
- Extension modules: Includes Digital Human (AI avatar lip-sync), Image-to-Video animation, and Motion Transfer (apply reference video motion to a target image).
- Available now: Apache 2.0 license; Windows one-click package available; macOS/Linux via
uv run streamlit run web/app.py.
Pixelle-Video is an open-source, fully automated short video engine from AIDC-AI , the AI lab inside Alibaba's international commerce division. You give it a topic, it gives you a finished MP4. No timeline, no keyframes, no manual editing. The engine handles everything: writing the narration script, generating AI illustrations or video clips for each scene, synthesizing a voiceover, adding background music, and stitching it all together. It has crossed 7,200 GitHub stars, which is a strong signal that this fills a real gap.
The pain point it targets is obvious in hindsight. Existing batch video tools either cost $50+/month or produce output that looks like it was assembled by a template bot from 2019. Pixelle-Video sits in a different category: it is infrastructure you run yourself, built on top of ComfyUI and standard LLM APIs. That means you control costs, you control quality, and you can swap out every component.
One topic in, one video out
You type a subject like "Why develop a reading habit" and the engine automatically writes the narration script, generates matching AI illustrations or video clips, synthesizes voice-over audio, adds background music, and composes everything into a finished MP4 file. The whole flow is described in the README as: script generation → image planning → frame-by-frame processing → video composition. Each step is modular and replaceable.

The system is designed around a pipeline architecture where each stage of video production is handled by a dedicated service. The core service layer , PixelleVideoCore , orchestrates LLM text generation, TTS voice synthesis, media generation, frame rendering via HTML templates, and final video composition through ffmpeg.
Frames are rendered by converting HTML templates to video segments using Playwright, which is what gives the output a polished, designed look rather than a raw AI dump.
The architecture under the hood
The pipeline runs 8 steps end-to-end, each implemented as an async lifecycle method on a