Mago Studio launches local AI video production with 8 GB of VRAM

AI video production has an awkward practical problem: the script sits in one service, images in another, video is generated in a third, and successful versions have to be hunted down in folders with names such as final_v7_really_final. Mago Studio attempts to bring this pipeline together in a single desktop application. The project is stored locally, renders can run on your own GPU through ComfyUI, and individual scenes can be sent to the cloud when necessary.

This is not merely another wrapper with a prompt field. Mago Studio has two primary modes: Director turns a script into characters, a shot list, a storyboard, and connected scenes, while Transform edits existing footage by changing its style, lighting, characters, or individual masked areas. Versions are retained between attempts, allowing the result to be discussed and refined as a project instead of treated as an endless feed of random generations.

What this looks like in practice

Suppose a team is making a one-minute product-launch video. It first uploads the script and breaks it into shots. For each shot, the team defines characters, references, and the visual direction. The next scene can continue from the previous one, helping maintain at least basic consistency in characters and environments. If one segment fails, it can be regenerated or edited separately without rebuilding the entire video.

The local component runs on top of ComfyUI. Windows users get an installer that sets up a compatible ComfyUI version, the required nodes, and models. The minimum configuration includes MiniMax H3 for video and Qwen3-VL 4B for describing images and videos. The developers specify an NVIDIA GPU with 8 GB of VRAM and approximately 63 GB of free storage. On macOS and Linux, the application connects to an existing ComfyUI installation. If no local GPU is available, the interface can still be used and renders can be run through Mago Cloud.

Cloud mode means the product is not entirely local: cloud rendering sends the selected shot, source video, references, and masks. According to the developers, the rest of the project remains on the machine. Scripts and reference images also use cloud models by default after account sign-in; fully offline operation requires separately installing a local LLM and image models.

Why open source matters here

Mago Studio is released under AGPL-3.0. This makes it possible to inspect how the application works, connect custom models, and adapt the workflow to a team's needs. Particularly useful is the ability to avoid waiting for the owner of a closed service to add a required video model: teams can implement the integration themselves through the existing ComfyUI pipeline.

AGPL, however, is not unconditional permission to do anything with the code. If a company modifies the application and makes it available to users over a network, it should review the source-disclosure obligations in advance. Licenses for downloaded models also apply separately: MiniMax H3, for example, is distributed under its own terms.

What to check before adoption

The project is still young, so an attractive demo video is no substitute for testing it in a real production workflow. Take one short script, run it from storyboarding through export, and measure the time required, VRAM usage, character consistency, and number of manual corrections. Test the ComfyUI setup separately as well: mismatched node and Python versions remain a common source of failures.

For a small studio or product team, Mago Studio is compelling not because it promises a magic button, but because it attempts to bring order to the process. If the tool survives your pilot project, it could replace a sprawl of tabs, folders, and spreadsheets with a single manageable pipeline. If it does not, the open-source code still gives developers a starting point for building their own local video-production system.