MoneyPrinterTurbo Guide: How It Works and How to Make AI Shorts Worth Watching

MoneyPrinterTurbo is a free, open-source tool that turns a topic or script into a finished short video with an AI voice, subtitles, footage and music. Here is how it works, how to install it, and what we learned making real Shorts with it, including why we pulled our Arabic versions.

By Quantara Editorial Team · Content Creation · Published Sep 29, 2026 · 9 min read

Short answer: MoneyPrinterTurbo is a free, MIT-licensed Python app that turns a topic or a finished script into a vertical short video. It writes or takes a script, finds footage, generates an AI voiceover, times subtitles to the voice, adds music and renders an MP4. It runs on an ordinary laptop. It is excellent at assembly and weak at ideas: the quality of the result depends almost entirely on the script and the visuals you give it.

How MoneyPrinterTurbo works

The pipeline has six stages, and you can stop after any of them:

1. Script. An LLM writes narration from a topic, or you supply your own. It supports OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Qwen, local models through Ollama, and many gateways.
2. Search terms. The LLM turns the script into visual keywords, or you supply them.
3. Footage. Those terms search free stock libraries (Pexels, Pixabay), paid ones, AI video models, or your own local images and clips.
4. Voice. Text-to-speech reads the script. Microsoft Edge TTS is free and needs no API key, and includes English and Arabic voices.
5. Subtitles. Captions are timed from the TTS word timings, or transcribed with Whisper for better accuracy.
6. Render. Clips, voice, subtitles and background music are assembled into a 9:16, 16:9 or 1:1 MP4.

You can drive it from a browser interface, a REST API, a command line with batch files of up to 100 videos, or Docker. The project is very popular, with about 127,000 GitHub stars at the time of writing.

What we made with it

To test it on a real brief, we produced five English Shorts: three about Verbito's Token Saver and two about the free Prompt Scorer. Instead of stock footage, the visuals are real screens from verbito.ai and branded explainer cards, and every number on screen is measured or a real live result, not an estimate.

The Token Saver Shorts are on the AI Token Saver page and the Prompt Scorer Shorts on the Prompt Scorer page.

Installing it

You need Python 3.11 or later. The fastest path on macOS or Linux uses uv:

git clone https://github.com/harry0703/MoneyPrinterTurbo.git
cd MoneyPrinterTurbo
uv sync --python 3.11
cp config.example.toml config.toml

FFmpeg is downloaded automatically through the imageio-ffmpeg package, so you do not need to install it separately. If you prefer containers, the project publishes a prebuilt Docker image. For generated scripts or stock footage, add an LLM key and a free Pexels key to config.toml. If you bring your own script and footage, as we did, you need neither.

Running it with your own script and footage

The command line lets you skip the parts you want to control:

.venv/bin/python cli.py \
  --video-script "Every time you ask ChatGPT something, you pay for the small talk..." \
  --video-terms "verbito" \
  --video-source local \
  --video-materials plate.png,card1.png,screen1.png \
  --video-concat-mode sequential \
  --voice-name en-US-AndrewNeural-Male \
  --subtitle-enabled --video-aspect 9:16

Supplying --video-script skips LLM script writing, supplying --video-terms skips keyword generation, and --video-source local uses your own images and clips.

Six things that tripped us up

1. A custom voice track silently removes subtitles

We rendered the voice once, then passed it back with --custom-audio-file to keep timing stable. The videos came out with no subtitles and no error. With a custom audio file there are no Edge TTS word timings, and the default subtitle provider skips quietly. Either let MoneyPrinterTurbo run TTS in the same job (the same text and voice produce the same audio), or switch the subtitle provider to Whisper.

2. Still images in equal slots look like a slideshow

By default every image gets the same duration and a hard cut, so a card about "Ultra" can appear while the voice is still on "Lite", and nothing moves. Our first cut fixed the timing by repeating frames in one-second slots, and it still looked like a slideshow. The version we published pre-renders the whole visual track with FFmpeg: each frame gets a slow, eased Ken Burns push-in or drift, frames cross-fade exactly at the subtitle cue where their topic is spoken, and a thin progress bar runs along the bottom. That single clip goes to MoneyPrinterTurbo as its only local material, and MoneyPrinterTurbo adds the voice, word-by-word animated captions (--subtitle-display-mode word_by_word --subtitle-animation pop_spring) and music. It is also much faster to render than dozens of one-second clips.

3. Arabic subtitles need proper text shaping

MoneyPrinterTurbo draws subtitles with Pillow. On our machine Pillow was built without libraqm, which would have rendered Arabic with disconnected letters in the wrong order. We generated the Arabic videos without subtitles and burned the timed subtitle file with FFmpeg's ass filter and shaping=complex, using an Arabic font (Cairo). Check an Arabic frame before you post; broken Arabic is the fastest way to lose an Arabic-speaking audience.

4. Stock footage makes product videos generic

For a software product, clips of people typing on laptops say nothing. We fed real screenshots of the product (the Token Saver selector, a real generated prompt, a real quality score) plus explainer cards as local materials. Viewers can see the feature working, which stock footage cannot show.

5. Put only real numbers on screen

Our first draft reused estimated token counts from our own page. Before rendering, we measured them with OpenAI's o200k tokenizer and found the estimates were off. The caveman prompting example is 165 tokens as a normal answer and 57 as a caveman answer, so we corrected the page and the videos. Short videos get screenshotted and checked; one wrong number costs more than it gains.

6. Free Arabic voices mispronounce colloquial scripts

We also made Egyptian Arabic versions with the free Edge TTS voices, and pulled them: native listeners heard the pronunciation as wrong. The voices read undiacritized Egyptian colloquial text with standard-Arabic habits. Rewriting in Modern Standard Arabic with full diacritics did not measurably help either; a Whisper transcription round trip actually matched the colloquial version more closely. For Arabic Shorts that sound native, use a human voiceover or a paid voice with a proper Egyptian or Gulf accent, and pass it to MoneyPrinterTurbo as custom audio (remembering lesson 1 about subtitles). Arabic subtitles, by contrast, are solvable for free, as lesson 3 shows.

Will AI-made Shorts actually get views?

Only if each one says something specific. YouTube, TikTok and Instagram demote repetitive, mass-produced content, and YouTube does not monetize content it considers inauthentic. Use automation for the assembly work (voice, timing, subtitles, rendering), and put your effort into the hook, the script and visuals that show something real. Label AI-generated content where a platform asks.

Get the script right first

The single biggest quality lever is the script: a hook in the first three seconds, narration sized to the video length (about 70 to 80 words for 30 seconds), one idea per beat, and concrete search terms for the visuals. Verbito's free AI Shorts Script Generator builds that prompt for you, in a format you can paste into MoneyPrinterTurbo's --video-script and --video-terms.