Pixelle-Video: type a topic, get a finished short video

Turning a topic into a short video usually means writing a script, sourcing visuals, recording a voiceover and editing — half a day gone by hand. Pixelle-Video, open-sourced by Alibaba’s AIDC-AI team, collapses that pipeline into one input box: hand it a theme like “why reading is worth the habit”, and it writes the script, generates an illustration or video clip for each scene, synthesizes the narration, and renders a finished cut with subtitles and background music. It is the “open-source fully automatic video editor” the WeChat Channels account 赛博煎蛋 featured on September 26, and the repo had 28,572 stars as of October 2, 2026.

Core features

• Topic to finished video: the pipeline runs script generation → shot planning → per-frame rendering → final assembly, and the README’s own UI screenshot shows a 10-scene 1080×1920 video taking 234 seconds. Already have a script? The fixed-copy mode skips the AI writing step.
• AI-generated visuals, not scraped stock: images render by default through a local ComfyUI FLUX workflow, or you can call DashScope, OpenAI, Seedream, Seedance and Kling image or video APIs directly, with style pinned by a prompt prefix. Unlike MoneyPrinterTurbo, which pulls stock footage, nothing in the frame carries a licensing question.
• Voice and cloning: Edge-TTS works out of the box for free; Index-TTS accepts a reference clip for voice cloning; the TTS step is itself a swappable ComfyUI workflow.
• Templates are layouts: static, image and video HTML templates control the frame across portrait, landscape and square sizes — plain HTML files, so you can write your own.
• Extension pipelines: digital-human presenting, image-to-video and motion transfer ship as extension modules — upload a reference video and photo to transfer motion.

The whole flow, from text to rendered scenes:

Typical use cases

• Solo creators running knowledge or marketing channels: no camera, no editing skills — feed it topics and batch out videos.
• Ops teams with copy and their own footage: the fixed-copy mode plus the custom-footage pipeline analyzes your photos and videos and builds the script around them.
• Developers wiring up content automation: image, video, TTS and VLM are swappable atoms, and workflows and templates are plain files you can version or re-vendor.

Quick start

On Windows, download the all-in-one bundle from Releases, unzip and run — no Python, uv or ffmpeg needed; fill in API keys in the system settings on first launch. On macOS or Linux, install uv and ffmpeg, then:

git clone https://github.com/ATH-MaaS/Pixelle-Video.git
cd Pixelle-Video
uv run streamlit run web/app.py

Open the WebUI, pick an LLM (Qwen, GPT, DeepSeek or a local Ollama), configure image generation (local ComfyUI or a direct API), type a topic and hit Generate. The README’s fully free setup is Ollama plus a local ComfyUI — zero cost.

Summary

If you want to batch-produce short videos and can live with AI-illustration looks, try this before paying for anything: unlike the earlier Story-Flicks, visuals can extend to video clips, digital humans and motion transfer. Already have footage and just want highlights cut automatically? That is AutoClip.

Three caveats: the free local tier needs an NVIDIA GPU for ComfyUI, so without one you pay for Qwen plus cloud rendering; upstream activity has slowed — the last commit was June 14, 2026, the newest release is the v0.1.15 Windows bundle from January 2026, and the GitHub org was renamed from AIDC-AI to ATH-MaaS (old links redirect); and reviews warn that AI illustrations converge on the same look, while purely AI-made content can get throttled on some platforms. Apache-2.0, Python.