Viggle Open-Sources Remake Weights: Swap the Actor in Finished Video via LoRA and ComfyUI

Anyone who makes video knows the pain: the clip is shot, edited and approved, and suddenly it turns out the actor doesn't fit, the logo on the table is wrong, and the lighting should have been evening. Reshooting is expensive. Generating a clip from scratch is not an option, because you would lose the very timing and camera movement the shoot was meant to capture. Viggle offers a third path and has released the weights for Remake, a model for editing finished video.

What Remake Can Actually Do

The tool has two modes. The first is called Recast: it replaces an actor or character based on a reference image. The second changes objects, environment or lighting. In both cases the source clip serves as the foundation, which means motion, timing and camera trajectory stay the same. The model does not invent the scene anew; it redraws the necessary elements on top of the existing structure.

Imagine a commercial in which a person walks down the street with a coffee. With Recast you drop in a different character, for example a brand mascot, while the gait and camera movement are preserved. With the scene mode you can turn a daytime street into an evening one or swap a cup for a bottle. The editing cuts do not need to be rebuilt.

What's Inside and How Much It Weighs

The release is built on MiniMax H3. Viggle distributes two separate LoRA adapters, one for each mode, along with a launch script, ComfyUI graphs and quantized variants. For those used to assembling pipelines from nodes, this is welcome news: ready-made graphs save days of tuning settings.

The price of openness is size. The base transformer at full precision takes up about 144 GB. This is not a model you run on a gaming laptop without preparation. That is why the documentation separately describes int8 and GGUF variants and offloading weights to system RAM. Speed and memory consumption depend on the configuration and the length of the clip, so nobody promises exact figures in advance. Plan tests on your own hardware rather than trusting other people's benchmarks.

An MCP Server for Long Videos

The most practical detail for agent developers is the MCP server. An agent can split a video into shots, process each one separately and assemble the result back together. Previously, character replacement across a series of shots had to be configured manually for each one. Now it can be described as a task for an agent: "replace the hero with the reference throughout the whole clip."

This is a typical pattern: a heavy generative model plus a thin orchestration layer that hides manual routine. If you are building a video service, the MCP wrapper shows how to connect such a model to an agent pipeline without custom-written stitching.

Limitations and What to Do

A few things deserve a sober look. First, the material contains no quality metrics: how stably the face or clothing holds up over long shots is unknown. Second, 144 GB at full precision means production will require serious hardware or rented GPUs. Third, quantization usually costs something in quality, and how much needs to be checked on your own clips. Finally, replacing people in video raises questions of likeness rights and labeling of synthetic content, so it is better to bring in lawyers early.

The practical takeaway is this. If you build tools for editing, advertising or content, download the LoRA adapters and ComfyUI graphs and run a couple of your own clips on int8 or GGUF. Product managers should assess whether reshoots can be replaced by editing finished material. Founders should note that open weights lower the barrier to entry for competitors who previously depended on closed APIs. And those building agents will find it useful to look at the MCP server as a ready-made example of working with video by shots.