Viggle Remake edits video locally: ComfyUI needs 12–24 GB of VRAM

An ad is finished, but the client wants a different character and a winter backdrop. That usually means a reshoot or working on each shot individually. Viggle Remake offers a way to edit the original video: specify the change, prepare reference images and generate a new version of the scene. For a team building video production tools, it is an opportunity to test local processing on its own footage.

Viggle has published the weights, inference code and ComfyUI integration. However, the promise that it “runs on a home GPU” needs clarification: a manual workflow on a GPU with 12–24 GB of VRAM and autonomous processing by an agent require different configurations. The project documentation explicitly distinguishes between them.

What changes in the video

Remake is based on MiniMax-H3 with two specialized LoRA modules. The swap mode handles character replacement, while edit handles objects, surroundings, weather and lighting. Inputs consist of the original clip, an instruction and one to four prepared images. These define the appearance of the result and can be tied to specific frames.

Imagine a clothing retailer that needs variations of the same advertising shot of a model walking. The team prepares images featuring a new jacket and checks how the change carries through the movement. The practical question is straightforward: how many frames will need fixing after generation, and is that cheaper than a reshoot? An impressive example in a repository does not answer that on its own.

The developers describe preservation of motion, camera movement and the original audio, but also list failures: artifacts in fine details, problems with nonhuman characters in long shots and temporal shifts between scenes under certain settings. These materials do not yet justify promising seamless replacement of any actor. The limitations are published in the model card.

What the agent handles

The project includes an MCP server, an interface through which an agent accesses processing tools. According to the authors, the agent can detect cuts, prepare keyframes, process individual segments, inspect the result and assemble the video with its audio track. Connection instructions are provided in the MCP guide.

For a product, this is more interesting than a standalone “replace face” button: it opens up the possibility of building a task queue and connecting asset preparation to rendering. But the presence of MCP does not establish how reliably the entire process runs without human intervention. That needs to be measured on your own clips.

I would start with a single short shot in which the face, hands and object being changed are clearly visible. Then I would test a head turn, an occluded object and a transition between scenes. This set of tests will reveal the tool’s limits faster than trying to remake a minute-long ad straight away. Repeat generations should be counted separately: they can eat into the expected savings.

What running locally costs

Quantized versions and manual ComfyUI workflows are available for GPUs with 12–24 GB of VRAM. An autonomous agent on a single RTX 5090 with 32 GB requires a different setup; the developers specify around 60 GB of system RAM. In their measurement, processing a 345-frame window at 1280×736 takes approximately 28 minutes. This is a reference figure from the authors, rather than an independent test. Details are provided in the ComfyUI instructions.

For an interactive editor, then, the first step is to check whether the wait is acceptable. Speed requirements may be less demanding for overnight batch generation of variations. Developers should define success in advance: whether the character remains recognizable, the product is preserved, noticeable defects are absent and how much manual cleanup is required.

Before integrating it into a commercial product, the licenses also need to be considered separately: the code is published under Apache 2.0, while the weights are governed by the MiniMax H3 Community License. A local installation does not, by itself, mean unrestricted use. The terms are set out in the licensing section.

Remake’s practical value lies in the ability to build and measure your own video editing workflow. A sensible next step for a team is a small pilot that tracks time, repeat attempts and manual corrections. Those figures will show where the tool actually shortens production.