What if you could take an existing video, replace the actors, change the environment, or completely redesign the scene while keeping the original performance? And what if you could run the entire process on your own GPU?
Meet Viggle Remake, a new open-weight AI video editing system built on MiniMax H3.
Viggle has released its model adapters, inference code, ComfyUI workflows, quantized models, and even an MCP server for AI coding agents. And this is a pretty interesting development for AI-assisted filmmaking.
Instead of generating a completely new video from scratch, Viggle Remake starts with existing footage.
The model attempts to preserve the original camera movement, timing, character actions, and soundtrack while changing whatever you specify.
Two specialized AI models handle the work.
1. Recast: Replace characters in existing videos
The swap mode lets you replace one or several people or characters using visual references.
Imagine taking a dance video and replacing the performers with entirely different characters.
The choreography remains.
The camera movement remains.
Even the original soundtrack stays.
2. Edit: Transform objects and environments
The edit mode handles everything else:
Change clothing and costumes
Replace or remove objects
Add new elements to a scene
Change backgrounds and locations
Modify lighting and weather
Transform the overall visual style
Both modes are implemented as separate LoRA adapters for MiniMax H3.
And here's where things get technically interesting.
Viggle didn't release another enormous foundation model.
Instead, the developers published two specialized LoRAs that extend the existing MiniMax H3 video model.
Each full-rank adapter is approximately 2.67 GB.
The original MiniMax H3 model is considerably larger, requiring roughly 144 GB of model files in the reference configuration.
That's not exactly friendly to consumer hardware.
But Viggle has prepared smaller versions.
The release includes:
INT8 models: Approximately 21 GB each
GGUF Q8: Approximately 22.3 GB
GGUF Q6: Approximately 17.6 GB
GGUF Q4: Approximately 14.3 GB
ComfyUI workflows: Ready-made graphs for character replacement and video editing
CPU offloading: Options for running on GPUs with significantly less memory
The documentation describes configurations down to 12 GB VRAM, although these require compromises in speed and substantial system memory.
So yes, you can experiment without renting an expensive GPU server.
Just don't expect instant results on a mid-range graphics card.
But here's my favorite part.
Viggle Remake also includes an MCP server for Claude Code, Codex, and other compatible AI agents.
And that changes the workflow considerably.
Instead of manually editing every shot, you can give an agent a longer video and describe what you want.
The agent can:
Analyze the footage and identify individual shots.
Divide the video into manageable segments.
Prepare reference frames and editing instructions.
Run the appropriate model for each segment.
Review the generated results.
Assemble everything into a finished video while preserving the original soundtrack.
The underlying model handles relatively short video windows, typically around 12 seconds.
The agent handles the longer sequence.
In one demonstration, Viggle recreated a 116-second music video containing 69 cuts.
That's a rather different proposition from generating isolated five-second clips.
And this is what I find most interesting.
We're watching AI video tools evolve from standalone generators into components of complete production pipelines.
Generate the assets.
Prepare the references.
Modify individual shots.
Check continuity.
Assemble the final edit.
And increasingly, let an AI agent coordinate the process.
Of course, there are limitations.
Complex non-human characters, fine details, and long uninterrupted shots can still produce artifacts or inconsistencies.
The model also cannot independently verify whether a person has consented to having their likeness used. Only use footage and identities you have permission to modify, and clearly label synthetic results.
And while the inference code is Apache 2.0, the model weights use the MiniMax H3 Community License, with additional conditions for commercial use.
But the direction is fascinating.
We're moving from AI that generates videos to AI that can help reshoot existing footage without bringing everyone back to the studio.
And now the tools behind that workflow are becoming available to independent developers.
That's a pretty significant step for open-weight AI filmmaking.
Download Viggle Remake weights and code:
https://huggingface.co/Viggle/Viggle-Remake
ComfyUI setup and ready-made workflows:
https://huggingface.co/Viggle/Viggle-Remake/blob/main/COMFYUI.md
Try Viggle Remake online:
https://viggle.ai/models/viggle-remake
#ViggleRemake #MiniMaxH3 #ComfyUI #AIVideoEditing #AICharacterSwap #OpenWeights #AIFilmmaking #AIAgents
YouTube / ABVX