Model or dataset
ATH-MaaS/Pixelle-Video avatar
ATH-MaaS/Pixelle-Video

Pixelle-Video: a topic-to-video pipeline built on ComfyUI, RunningHub and direct model APIs

AI Fully Automated Short Video Engine.

30,127 stars4,373 forksPythonApache-2.0

At a glance

What is it?
Pixelle-Video turns a single topic string into a narrated, illustrated short video through a modular pipeline. The repository ships a Windows bundle, a Streamlit UI and a FastAPI backend, and it assumes you already have model access somewhere.
Who is it for?
Adopt Pixelle-Video if you already run ComfyUI or hold API keys for DashScope, OpenAI, Seedream, Seedance or Kling, and you want the script, images, voice and BGM assembled for you. Do not adopt it if you expect a self-contained generator with no external model accounts, if you need macOS or Linux packaging rather than source installs, or if you cannot absorb per-video API cost.
Can I use it commercially?
Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 154 days ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • ATH-MaaS/Pixelle-Video30,127 ★ · +1,821
Star history
Repository2026-09-212026-11-15
ATH-MaaS/Pixelle-Video28.3k30.1k

1,821 Stars in the last 55 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The gap Pixelle-Video fills between a topic and a finished cut

Most AI video tooling solves one step. An image model draws a frame, a TTS engine reads a line, an editing library concatenates clips. The work of deciding what the script says, how many frames it needs, which line gets which picture, and where the music sits is still manual. Pixelle-Video targets that assembly layer. The README states the input is a single topic and the output is a complete video, with the pipeline described as script generation, image planning, frame-by-frame processing, then video composition.

The intended user is someone producing short-form video at volume without editing experience. The README repeats the phrase zero barrier, zero editing experience, and the feature list is organized around that: AI-written narration, one illustration per sentence, TTS voice-over, background music, and template-based visual styling. It is not a timeline editor and it does not expose keyframes. If you want frame-accurate control, this is the wrong shape of tool.

The repository also lists extension pipelines beyond the default flow: digital human presentation, image-to-video, and motion transfer, the last added on 2026-01-26 according to the changelog. Those are separate modules rather than variations of the main path.

How the pipeline is wired: atomic capabilities behind one orchestrator

The README describes the design as modular, and the repository layout supports that reading. There is a pixelle_video package, a web directory for the Streamlit interface, an api directory for the FastAPI service, a workflows directory, and a templates directory. The stated goal is that image, video, TTS and VLM capabilities can each be swapped independently.

Two backends sit behind those capabilities. The first is workflow execution through ComfyUI or RunningHub, which means the actual generation graph is a workflow file you can edit. The second is direct API calls. The 2026-06-01 changelog entry adds direct API media model configuration, letting you set the provider, Base URL and a proxy toggle from the WebUI. The feature list names DashScope, OpenAI, Seedream, Seedance and Kling as reachable services. Text models are similarly pluggable, with GPT, Qwen, DeepSeek and Ollama listed.

That split matters for cost and for control. A ComfyUI workflow gives you the graph but requires the hardware or a rented GPU. RunningHub is the hosted route, and the changelog notes configurable concurrency limits and support for 48G VRAM machines. Direct APIs remove the GPU requirement entirely but move the work to a vendor. The engine is the conductor, not the instrument.

Installing Pixelle-Video on Windows versus from source

The README recommends the Windows bundle for Windows users and says it needs no Python, uv or ffmpeg install. The steps are to download the latest one-click bundle from the releases page, extract it, run start.bat, and let the browser open http://localhost:8501. Configuration happens afterward in the system settings panel, where you enter the LLM API and image generation service.

For macOS, Linux, or anyone who wants to modify the code, the source route requires two prerequisites. The first is uv, which the README points to the official uv installation guide for, with uv --version as the verification step. The second is ffmpeg, installed per platform:

bash
# macOS
brew install ffmpeg

# Ubuntu / Debian
sudo apt update
sudo apt install ffmpeg

The project metadata confirms the runtime floor. The pyproject file sets requires-python to >=3.11, and the dependency list pins edge-tts to exactly 7.2.7, moviepy to 1.0.3, and pillow below 12. That edge-tts pin is not incidental: the 2025-12-10 changelog entry says the version was locked to fix TTS service instability.

There is also a container path. The docker-compose file defines an init service that copies config.example.yaml to config.yaml if the file is missing, then an api service that runs the FastAPI backend on port 8000. A USE_CN_MIRROR build argument switches apt and pip to mirror sources, which the compose comments show as USE_CN_MIRROR=true docker-compose up -d. The Dockerfile installs ffmpeg and fonts-noto-cjk, the latter for CJK character rendering, so container output is not limited to Latin text.

What the first run actually asks of you

The Windows flow in the README is five steps: extract the bundle, run start.bat, open the browser at port 8501, configure the LLM API and image service in system settings, then generate. The bundle ships dependencies, so the first real friction is credentials rather than installation.

Source and Docker users hit the same wall. The repository ships config.example.yaml, and the configuration surface can be read from the example file or the documentation site at the project homepage. The docker-compose init service exists precisely because mounting a missing config.yaml creates a directory instead of a file, which tells you the config file is the expected integration point.

Once configured, the README's own framing of the first run is that one topic keyword produces a video with no editing. What you should expect to see is a script, a set of generated stills or clips, a synthesized voice track, background music, and a composed file.

Where Pixelle-Video breaks down

The engine has no model of its own. Every image, clip and voice line comes from a service you configured, which means output quality is a function of your provider choices, not of the project. A weak image model produces a weak video, and the README's example gallery shows results across several backends including Qwen for image generation and cloned voices for narration.

Cost is unbounded by design. A per-sentence illustration policy means a longer script triggers proportionally more image calls, and if you enable AI video generation per frame, the bill scales again. For a fixed-length short this is manageable; for batch generation it is a line item you have to monitor outside the tool.

Dependency pinning cuts both ways. Locking edge-tts to 7.2.7 fixed a reported instability, but it also means you cannot move that package forward without diverging from the project. moviepy at 1.0.3 and pillow under 12 are similar constraints that will eventually collide with other packages in a shared environment.

The packaging story is uneven. There is a Windows one-click bundle, but no equivalent macOS or Linux bundle is mentioned; those users get source or Docker. And the recent GitHub releases are all Windows bundles, the latest being v0.1.15 on 2026-01-27, while the source version in pyproject is 0.2.0. Those two numbering tracks are not obviously the same release line.

Pixelle-Video against assembling the same pipeline yourself

The obvious alternative is not another single product but a hand-built chain: an LLM call for the script, an image API for the frames, edge-tts or another TTS for narration, and ffmpeg for the concat. That is roughly what Pixelle-Video does internally, and the difference is where the decisions live.

In a hand-built chain you own the prompt that turns a topic into a script, the rule that maps sentences to images, and the timing logic that aligns audio to frames. In Pixelle-Video those live in templates and workflows inside the repository, which you can edit but must first locate. The trade is setup speed against control: the project gives you a working default and a Streamlit UI in exchange for adopting its structure.

A second alternative is staying entirely inside ComfyUI. If your generation already happens there, you can build the video graph yourself and skip the orchestration layer. Pixelle-Video's value in that case is the script and planning logic plus the UI, not the generation. If you have no interest in automatic scripting, the ComfyUI-only route is less machinery.

A third is a hosted video generation service with a single API call. Those trade flexibility for simplicity, and they typically will not let you swap the TTS engine or the image model independently. Pixelle-Video's stated position is the opposite: atomic capabilities you can replace one at a time.

Licence, maintenance and the cost of upgrading

The project is Apache-2.0, declared in both the LICENSE file and the pyproject license field, and the repository carries a NOTICE file. Apache-2.0 permits commercial use and modification and includes a patent grant, but it also carries attribution and notice-retention obligations. If you redistribute a modified version, the NOTICE handling is the part to read carefully. This is a description of the licence terms, not legal advice; get counsel for a commercial redistribution.

On maintenance, the last push to the default branch was on 2026-06-14. The most recent release listed is v0.1.15 from 2026-01-27, a Windows bundle. The changelog's newest entry is 2026-06-01, adding direct API media model configuration. So the code has moved more recently than the tagged releases, and the release tags track the Windows packaging rather than the Python package version, which pyproject puts at 0.2.0.

Upgrade cost is concentrated in three places: the pinned dependencies, the config.yaml schema, and the workflow files under workflows/ and templates/. A dependency bump that touches edge-tts or moviepy is a real change, not a version string edit. The docker-compose setup mitigates some of this by rebuilding the image, but the mounted config.yaml is yours to migrate by hand.

Editorial conclusion

Adopt Pixelle-Video if you already run ComfyUI or hold API keys for DashScope, OpenAI, Seedream, Seedance or Kling, and you want the script, images, voice and BGM assembled for you. Do not adopt it if you expect a self-contained generator with no external model accounts, if you need macOS or Linux packaging rather than source installs, or if you cannot absorb per-video API cost. Verify first that a single topic run completes end to end on your own keys and that the resulting video passes your quality bar before you build a workflow around it.

Frequently asked questions

What is Pixelle-Video and what does it generate?

It is an AI short video engine that takes a single topic as input and produces a finished video. According to the README, it writes the script, generates the illustrations or clips, synthesizes the voice-over, adds background music and composes the final file.

How do I install Pixelle-Video?

Windows users are pointed to the one-click bundle on the releases page, which the README says needs no Python, uv or ffmpeg install; you extract it and run start.bat. macOS and Linux users install uv and ffmpeg first, then set up from source. A docker-compose path also exists, running the FastAPI backend on port 8000.

Does Pixelle-Video work with ComfyUI?

Yes. The README lists ComfyUI and RunningHub workflows as one of the two generation backends, alongside direct API calls to services such as DashScope, OpenAI, Seedream, Seedance and Kling. The repository includes a workflows directory for those graphs.

Is Pixelle-Video free to use?

The software is Apache-2.0 licensed, so the code itself is free to use and modify. The generation is not: the engine calls external LLM, image, video and TTS services that you configure with your own credentials, and those providers bill you separately.

Official sources

  1. ATH-MaaS/Pixelle-Video on GitHub
  2. License: Apache-2.0
  3. Project website
  4. README
  5. Releases
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/ath-maas-pixelle-video.svg)](https://hysenlabs.com/projects/ath-maas-pixelle-video)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/ath-maas-pixelle-video.svg)](https://hysenlabs.com/projects/ath-maas-pixelle-video)

Was this useful?