Overview
In AI video generation, most tools specialize: image generation (Stable Diffusion), video generation (WAN), or speech synthesis — never all three. Pixelle-Video’s difference is that it integrates the whole video production chain into one fully automated, end-to-end pipeline. Input a topic and it produces a complete short video — copywriting, illustrations, voiceover, background music, and subtitles — in about three minutes.
The project comes from the Chinese development team AIDC-AI and has passed 13,600 stars on GitHub as a completely free open-source solution (Apache-2.0). Repository: github.com/AIDC-AI/Pixelle-Video. Documentation: aidc-ai.github.io/Pixelle-Video.
In One Sentence
Pixelle-Video = a fully automatic short-video engine. Topic in, finished video out. Its goal is to make video creation a one-sentence affair — no editing experience, no scriptwriting, no complex tooling to learn.
Core Capability: The Fully Automated Pipeline
The user supplies only a topic (say, “why you should build a reading habit”), and Pixelle-Video completes every step automatically:
- AI writes the script — generating the storyboard and narration from the topic
- AI generates the illustrations — a matching image for each scene
- AI synthesizes the voiceover — turning narration into natural speech
- Background music is added — with automatic BGM selection
- One-click final composition — all assets merged into the finished video
The Component Capability Matrix
| Stage | Options | Notes |
|---|---|---|
| LLM (copywriting) | GPT / Qwen / DeepSeek / Ollama | Smart narration generation with model switching |
| Image generation | ComfyUI (FLUX / Qwen etc.) / RunningHub | One matched illustration per sentence |
| Video generation | WAN 2.1 / WAN 2.2 / Nano Banana | AI-generated motion video as backgrounds |
| Speech synthesis | Edge-TTS / Index-TTS / ChatTTS / voice cloning | Multi-language, multi-voice, reference-audio cloning |
| Video composition | Built-in FFmpeg pipeline | Image-video mixing, background overlay, subtitles |
Architecture: Every Capability Is a ComfyUI Workflow
Pixelle-Video’s central design idea: all AI capabilities are defined through ComfyUI workflow JSON files. The pipeline is genuinely decoupled from specific models — to switch image generation from FLUX to Qwen, you don’t change code; you point to another workflow JSON. The layered architecture:
- Application layer — the Pixelle-Video web UI (Streamlit), the user-facing surface
- Orchestration layer — the video pipeline sequencing script → images → voice → composition
- Capability layer — ComfyUI workflow JSON files defining models and parameters
- Execution layer — ComfyUI (local or remote) running the workflows to produce images and video
This buys enormous flexibility: a better image model requires zero code changes — just add a new workflow JSON.
Feature Highlights
- Fully automatic generation — topic in, finished video out in about 3 minutes, no manual editing
- AI smart copywriting — narration and storyboards from any of several LLMs
- Flexible video dimensions — vertical (1080×1920), horizontal (1920×1080), and more
- Multiple visual styles — bundled templates for quick style switching
- Custom material — upload your own photos and videos; the AI recognizes content and writes descriptions
- Composable atomic capabilities — use preset ComfyUI workflows or author your own
- API support — a REST API for integration into other systems
Installation
Windows (recommended all-in-one package)
Download the latest Windows bundle from GitHub Releases, extract it, and double-click start.bat. No Python, FFmpeg, or dependency installation required.
macOS / Linux (from source)
git clone https://github.com/AIDC-AI/Pixelle-Video.git
cd Pixelle-Video
# uv recommended (a faster package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync
# start the web UI
uv run streamlit run web/app.pyOpen http://localhost:8501 in your browser. If you choose a locally deployed ComfyUI as the image backend, you’ll also need to install ComfyUI and download the model files.
Your First Video
Once the web UI is running, the workflow is simple: configure an LLM API key (Qwen, DeepSeek, or GPT — any one) and the image service address in system settings; enter a topic under content input, such as “the importance of healthy eating”; set the scene count (5 storyboards by default); choose the TTS and image-generation workflows; and hit the generate button. The system shows live progress — script, per-scene images, voice synthesis, final composition — and the whole run takes roughly two to five minutes.
What It Costs
Pixelle-Video scales to any budget. Completely free: Ollama for the LLM (local) plus a local ComfyUI deployment — zero cost, suited to users with an NVIDIA card (6GB+ VRAM). Recommended: Qwen for the LLM (about 0.01-0.05 RMB across three videos) plus local ComfyUI — outstanding value. Cloud: OpenAI API for the LLM plus RunningHub for images — higher cost but no local GPU needed.
Against Its Peers
Pixelle-Video didn’t start from zero — it borrows from several strong open-source projects while innovating in architecture: MoneyPrinterTurbo is the closest comparable tool, but Pixelle’s ComfyUI-workflow architecture wins on model flexibility and extensibility; NarratoAI focuses on film-commentary automation, where Pixelle-Video is more general-purpose; and Pixelle-MCP, from the same team, is the ComfyUI MCP server (one of Pixelle-Video’s underlying dependencies) that lets AI assistants drive ComfyUI directly.
Best-Fit Scenarios
- Content creators: batch-produce knowledge and explainers — topic in, video out
- Education and training: quick course-explainer and study-content videos
- Marketing: auto-generated product intros and brand-story material
- Personal use: travel vlogs and lifestyle clips for fun
Limits to Know
- Output quality is bounded by the underlying models (ComfyUI models, TTS quality)
- Local running has a configuration threshold (6GB+ VRAM recommended), though cloud options remove it
- 63 open issues on GitHub at review time, mostly configuration-dependency and subtitle-generation functionality
- The style-template library is still expanding
The Bottom Line
Pixelle-Video is among the most active AI short-video generation projects on GitHub, and its 13,600-plus stars reflect community conviction in the direction. The core idea — decoupling AI capabilities through ComfyUI workflows — is architecturally forward-looking: users can swap and upgrade underlying models without touching a line of code. For creators who need to batch-produce short videos, it is a genuinely practical free open-source solution. The low barrier (one-click Windows bundle), the zero cost (local deployment), and the fully automated pipeline lift video-creation efficiency by an order of magnitude.
Tags: Pixelle-Video, AI video generation, short video, open source, ComfyUI, AIGC, content creation
