---
title: "Pixelle-Video: One Topic, a ComfyUI Assembly Line"
description: "Streamlit app that turns a topic or script into short video: LLM copy, ComfyUI or RunningHub media, TTS, BGM, HTML templates, ffmpeg out to output/."
pubDate: "2026-05-07"
lang: "en"
slug: "pixelle-video-comfyui-short-form-engine"
url: "https://hoangyell.com/pixelle-video-comfyui-short-form-engine/"
author: "HoangYell"
translation_vi: "https://hoangyell.com/vi/pixelle-video-comfyui-short-form-engine/"
categories: ["AI & Agents"]
tags: ["comfyui", "short-video", "neural-audio", "streamlit", "aigc", "video-generation"]
heroImage: "https://raw.githubusercontent.com/yellcamap/img/refs/heads/main/photo-1485846234645-a62644f84728.webp"
---

## TL;DR

- Feeds a topic or a fixed script through script planning, per-shot media, voice, optional BGM, then ffmpeg composition into `output/`, all from a three-column Streamlit UI ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).
- Image and motion come from ComfyUI locally or from RunningHub JSON workflows; you swap graphs by dropping API-format files named `image_*`, `video_*`, or `tts_*` into `workflows/` ([the FAQ](https://github.com/AIDC-AI/Pixelle-Video/blob/main/docs/FAQ.md)).
- Best for makers who already speak ComfyUI and want templated vertical or landscape shorts without living inside a nonlinear editor ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).
- Differentiator: atomic workflows you can remix (FLUX instead of default image graph, ChatTTS instead of Edge) while the rest of the pipeline stays put ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).
- Concrete path: `uv run streamlit run web/app.py`, open `http://localhost:8501`, save LLM + ComfyUI or RunningHub keys, click generate, read the mp4 path the UI prints beside duration and shot count ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md), [the FAQ](https://github.com/AIDC-AI/Pixelle-Video/blob/main/docs/FAQ.md)).

The upload bar is not the enemy this time. The enemy is shot five of seven still waiting on a GPU queue while your caption writer pings you for a script tweak. I stared at a blank timeline and realized I was not stuck on creativity. I was stuck on handoffs.

## Situation: shorts without a human conveyor belt

Short-form video is a kitchen during rush hour. Someone writes the ticket, someone fires the line, someone plates. When you are all three roles, the metaphor breaks and you only feel the heat. Pixelle-Video is the expediter: it keeps the ticket readable, routes each dish to the right station, and brings back a finished plate in `output/` ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)). The tension that remains is whether your ComfyUI stations stay online.

## Task: what it does in one sentence

This repo is Pixelle-Video, an Apache-2.0 Streamlit app that automates short video creation by chaining an OpenAI-compatible LLM (optional), ComfyUI or RunningHub media workflows, TTS, optional background music, and HTML templates, then ffmpeg, so a topic or script becomes a rendered file without manual NLE steps ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).

## Real-world use cases

- **Explainer shorts from a single prompt**: AI Generated Content mode turns a topic like "why read daily" into narration plus visuals ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).
- **Narration-first with your own words**: Fixed Script Content skips the LLM writer and still generates frames and voice ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).
- **Brand voice via reference audio**: upload MP3, WAV, or FLAC for cloning workflows such as Index-TTS graphs you place under `tts_` prefixes ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).
- **Template A/B for aspect ratio**: pick `static_*`, `image_*`, or `video_*` HTML templates grouped by portrait, square, or landscape ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).
- **Custom Comfy graphs**: bind `$prompt.text!` or `$prompt.value!` on the CLIP text node title, export API Format, name files correctly, reload the UI ([the FAQ](https://github.com/AIDC-AI/Pixelle-Video/blob/main/docs/FAQ.md)).

Walkthrough: before and after the button

```
# Before (manual NLE mental load)
1. Draft script in a doc.
2. Generate stills or clips in separate tools.
3. Record VO or run TTS elsewhere.
4. Align layers, levels, and exports in an editor.
```

```
# After (Pixelle-Video happy path)
git clone https://github.com/AIDC-AI/Pixelle-Video.git
cd Pixelle-Video
uv run streamlit run web/app.py
# Browser opens http://localhost:8501
# System Configuration → LLM + ComfyUI http://127.0.0.1:8188 or RunningHub key → Save
# Generate Video → watch progress → open output/<generated>.mp4
```

Expected observation: the UI shows steps such as "Frame 3/5 - Generating Image" and ends with duration, size, and shot count next to the file ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md), [the FAQ](https://github.com/AIDC-AI/Pixelle-Video/blob/main/docs/FAQ.md)). But what happens when the free TTS lane stalls?

## How to use it

Top commands and what you should see:

1. **Clone and enter the tree**
   - Command: `git clone https://github.com/AIDC-AI/Pixelle-Video.git && cd Pixelle-Video`
   - Outcome: repository with `web/app.py`, `workflows/`, `templates/`, `bgm/` ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).

2. **Launch the Streamlit shell (uv path)**
   - Command: `uv run streamlit run web/app.py`
   - Outcome: local server on port `8501`, browser tab with three columns ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).

3. **Windows all-in-one**
   - Action: download the latest release bundle, run `start.bat`
   - Outcome: same port without installing Python manually ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).

4. **Configure media backends**
   - Local: set ComfyUI URL (default `http://127.0.0.1:8188`) and hit Test Connection
   - Cloud: paste RunningHub API key when you use bundled `runninghub/*.json` workflows ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md), `config.example.yaml`).

5. **Verify artifacts**
   - Command: `ls output/` after a successful run
   - Outcome: composed mp4 files the UI referenced ([the FAQ](https://github.com/AIDC-AI/Pixelle-Video/blob/main/docs/FAQ.md)).

Directory shape worth memorizing:

```text
Pixelle-Video/
  web/app.py          # Streamlit entry
  workflows/          # image_*.json, video_*.json, tts_*.json (API format)
  templates/          # static_*, image_*, video_* HTML
  bgm/                # drop custom MP3/WAV
  output/             # rendered videos land here
```

Markdy scene: the night the TTS lane caught fire

## Configuration and customization

- **LLM block in `config.example.yaml`**: set `api_key`, `base_url`, and `model` for any OpenAI-compatible endpoint. Change when you outgrow the bundled presets (Qwen, GPT, DeepSeek, Ollama URLs are commented inline) ([`config.example.yaml`](https://github.com/AIDC-AI/Pixelle-Video/blob/main/config.example.yaml)).
- **ComfyUI URL vs Docker**: use `http://127.0.0.1:8188` on bare metal; switch to `host.docker.internal:8188` on Mac or Windows containers, or your host IP on Linux, so queues still reach the GPU box ([`config.example.yaml`](https://github.com/AIDC-AI/Pixelle-Video/blob/main/config.example.yaml)).
- **`runninghub_concurrent_limit`**: raise only when your membership tier allows more parallel cloud jobs; keeps bills predictable ([`config.example.yaml`](https://github.com/AIDC-AI/Pixelle-Video/blob/main/config.example.yaml)).
- **`prompt_prefix`**: English style string applied before every image or video prompt; tighten it when brand guidelines demand a single look ([`config.example.yaml`](https://github.com/AIDC-AI/Pixelle-Video/blob/main/config.example.yaml)).
- **`template.default_template`**: points at paths such as `1080x1920/image_default.html`; change when your distribution channel demands a new aspect ratio ([`config.example.yaml`](https://github.com/AIDC-AI/Pixelle-Video/blob/main/config.example.yaml)).

💡 **Tip:** Copy `config.example.yaml` to `config.yaml` for static defaults, but remember the UI also writes runtime settings. Never commit secrets ([`config.example.yaml`](https://github.com/AIDC-AI/Pixelle-Video/blob/main/config.example.yaml)).

## Where it fits (and where it does not)

- Fits when ComfyUI or RunningHub is already part of your stack and you want a guided short-video wrapper with history, batch tasks, and motion-transfer add-ons noted in recent release bullets ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).
- Does not fit when you need a fully managed editor with collaborative timelines and manual keyframes; this is automation-first.
- Works alongside Pixelle-MCP if you also want agents invoking ComfyUI directly ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).

## The rough edges

- Edge-TTS depends on Microsoft’s free surface and may fail on noisy networks; the FAQ explicitly recommends ComfyUI `tts_` workflows for stability ([the FAQ](https://github.com/AIDC-AI/Pixelle-Video/blob/main/docs/FAQ.md)).
- LLM failures usually trace to bad `base_url`, exhausted balances, or typoed model names; the FAQ tells you to verify each ([the FAQ](https://github.com/AIDC-AI/Pixelle-Video/blob/main/docs/FAQ.md)).
- Features that rely on Chromium may throw "Could not find a Chrome executable"; installing Chrome resolves it ([the FAQ](https://github.com/AIDC-AI/Pixelle-Video/blob/main/docs/FAQ.md)).
- Generation latency grows with shot count, model speed, and WAN conditions; README sets expectations in minutes, not seconds ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)).

⚠️ **Warning:** Treat API keys like cash. Streamlit binds locally, yet any script with your RunningHub or LLM tokens can rack up charges if you expose the port.

## How it compares: alternatives the repo names

The README credits [Pixelle-MCP](https://github.com/AIDC-AI/Pixelle-MCP), [MoneyPrinterTurbo](https://github.com/harry0703/MoneyPrinterTurbo), [NarratoAI](https://github.com/linyqh/NarratoAI), [MoneyPrinterPlus](https://github.com/ddean2009/MoneyPrinterPlus), and [ComfyKit](https://github.com/AIDC-AI/ComfyKit) as inspiration ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)). Pixelle-Video’s bet is composable Comfy graphs plus a Streamlit control surface instead of a monolithic editor.

## Final thoughts

I stopped measuring the night by how many keyframes I touched. The concrete result is a rendered mp4 in `output/`, a UI that narrates each shot, and a community that pushed the repo past twelve thousand stars while shipping motion transfer and digital-human pipelines ([the README](https://github.com/AIDC-AI/Pixelle-Video/blob/main/README_EN.md)). The kitchen still gets hot, but the tickets finally print legibly.

[AIDC-AI/Pixelle-Video](https://github.com/AIDC-AI/Pixelle-Video) · Apache-2.0 · 12738 · [docs](https://aidc-ai.github.io/Pixelle-Video/zh)