Maestro: video generator — what it does and what you need to run it

Maestro is an open-source video generator from Blizaine, positioned as a local studio for creating video, images and music. It works on the image-text-to-video scheme: text and optionally an image go in, and a video sequence comes out. Installation is done through the Pinokio platform.

What it does

As described by the authors, Maestro combines video, image and music generation in a single tool. Director mode lets you plan full-fledged music videos and short films from a single text prompt. Video generation is based on the WanGP pipeline. The language model uses the minimax-h3 library, and the model is downloaded from huggingface.co/MiniMaxAI/MiniMax-H3.

The authors note that Maestro automatically downloads llama-server (about 600 MB, one-time) and the selected GGUF model on first launch. Gemma 4 4B is offered by default — according to the developer, it runs fast and comfortably even on small GPUs. The system detects CUDA on its own and binds the language model to the graphics card if one is available.

What you need to run it

The platform, as described by the author, is CUDA. VRAM requirements: 12—24 GB, this is the developer's claim, not the result of independent measurements. The description mentions that after an update, cards with 12 GB and more get a fast language model engine; previously the audio profile mistakenly blocked it on cards with less than 24 GB of memory.

The Python version, as described by the author, is 3.11. The developer states that the verified environment of Python 3.11 / PyTorch 2.10 / CUDA 13 has become standard for installation, updates and running on compatible systems with RTX 40- and 50-series graphics cards, while fallback options for other configurations are preserved.

The largest model weights file takes up 9.7 GB, and the total size of all model weights files is 464.1 GB. The code is distributed under the WanGP Non-Commercial Evaluation License 1.1, which permits non-commercial use only. The weights are distributed under the MiniMax H3 Community License, which permits commercial use with annual revenue of up to 20 million USD. The code is written in Python, and the repository is available at github.com/Blizaine/Maestro. The last code change was on September 4, 2026, and the repository was created on July 8, 2026.

Who it suits

The tool is aimed at those who need fully local video generation without resorting to cloud services. The presence of Director mode may be useful for authors who assemble music videos or short films from a single script. The open source allows you to understand how the pipeline works and, if needed, adapt it to your own tasks.

It is not suitable for those who do not have a CUDA-capable graphics card with enough VRAM: as described by the author, comfortable operation requires at least 12 GB. The total size of the weights at 464.1 GB requires considerable disk space if you download all the files. The code under the WanGP Non-Commercial Evaluation License 1.1 may be used for non-commercial purposes only, while the weights under the MiniMax H3 Community License may be used commercially with annual revenue of up to 20 million USD.

Maestro is a local open-source video generator that combines several types of generation in a single application. It is aimed at users with CUDA-compatible graphics cards and enough memory, and installation through Pinokio simplifies deployment. The code is distributed under the WanGP Non-Commercial Evaluation License 1.1, and the weights under the MiniMax H3 Community License.