Overview

Viggle Remake is a video-to-video editing model built from rank-256 LoRA adapters on MiniMax-H3. It supports character replacement as well as broader edits to objects, backgrounds, weather, lighting, and style. The maintainer is Viggle.

Best use cases

Use it to recast people or characters, change outfits, add or remove objects, and alter a video's setting, weather, lighting, or style while retaining source motion, camera, timing, and soundtrack. The repository describes both single-shot use and workflows that split longer films into windows.

Limitations

The reference sampler handles one window of about 12 seconds; longer videos require windowing, which the MCP workflow can perform. The base MiniMax-H3 files total about 144 GB, and the reference sampler requires the base text encoder and one GPU with at least 80 GB memory at bf16; ComfyUI alternatives are documented for smaller cards. A six-section prompt is required, and timestamps must fit the requested frame count. Painted reference frames determine identity and appearance; pinned frames also constrain composition. Audio is carried from the source rather than generated. Documented failure cases include non-humanoid characters in a long take, fine-detail artifacts, and timing drift in multi-shot clips rendered without pins. The model cannot verify identity or consent; use footage and likenesses only with appropriate rights and consent, and label generated content as AI-generated.

How it compares

Viggle-Animate is described as replacing a character, while Remake also supports object insertion or removal and changes to setting, weather, and lighting. Viggle-Animate is the supplied internal alternative. Other supplied alternatives include LTX-2 Retake, Luma Ray 3.2 Reframe, Ray2 Flash Modify, and Luma Modify Video; the supplied descriptions do not establish a quality, speed, or cost ranking. The README identifies Remake as the “Viggle-Animate v2” previewed on the Viggle-Animate card; that is a version relationship, not evidence that the separate Animate model has Remake's full feature set.

Technical specifications

The model consists of two PEFT LoRA adapters, swap and edit, each rank 256 and about 2.67 GB, applied to the MiniMax-H3 base transformer. Lower-rank variants at ranks 128 and 64 are also provided, at about 1.34 GB and 0.67 GB respectively; the README describes their output as near-identical. The required base files are about 144 GB. The repository also contains ComfyUI assets and merged int8 and GGUF transformers. The reference sampler uses bf16 and requires an 80 GB-or-larger GPU. It renders at 24 fps, with frame counts from 90 to 345 following 17n + 5; the README recommends up to 294 frames. At short edge 736, the documented output size is 1280×736. No training data, training procedure, or measured benchmark results are supplied. The weights are derivatives of MiniMax-H3 and are subject to the MiniMax H3 Community License; the supplied code directories are Apache 2.0.

Model inputs and outputs

Inputs

The reference sampler takes a source video, a prompt in MiniMax-H3's six-section format, and one to four painted reference images. Each reference can optionally be pinned to the source frame it was painted from. Select swap to replace people or characters, or edit for other edits such as object changes and background, weather, or lighting changes. The prompt's timestamps must fit the selected frame count. The sampler accepts a seed and short-edge setting; ffmpeg must be available on PATH.

Outputs

The sampler produces a rendered video window and carries over the source soundtrack; it does not generate audio, so silent input produces silent output. The MCP workflow can split a longer video into windows and splice the rendered results with the soundtrack. The README documents 24 fps output and a 1280×736 size when --short-edge 736 is used.

Getting started

hf download MiniMaxAI/MiniMax-H3 --local-dir ./MiniMax-H3 \ --include "modular_model_index.json" "transformer/*" "text_encoder/*" "tokenizer/*" \ "processor/*" "vae/*" "audio_vae/*" "scheduler/*" "audio_scheduler/*" "docs/*" hf download Viggle/Viggle-Remake --local-dir ./Viggle-Remake \ --include "lora/*" "inference/*" "examples/*" "requirements.txt" pip install -r Viggle-Remake/requirements.txt python Viggle-Remake/inference/sample.py --mode swap \ --video source.mp4 --prompt prompt.txt \ --ref ref1.png --ref ref2.png --pin 12 --pin 96 \ --frames 124 --short-edge 480 --seed 0 \ --base ./MiniMax-H3 --lora-repo ./Viggle-Remake --out out.mp4

The reference sampler also requires an ffmpeg binary on PATH and an 80 GB-or-larger GPU for bf16. The prompt must follow the six-section format, and the reference images must be painted frames.

Frequently asked questions

Q: What are the two modes?

A: swap replaces one or more people or characters. edit handles other changes, including object insertion or removal and changes to background, weather, or lighting.


Q: How many reference images can I provide?

A: The sampler accepts one to four painted reference images. Each can optionally be pinned to the source frame it corresponds to.


Q: Can it generate or change the soundtrack?

A: No. It carries the source soundtrack into the output; a silent source produces a silent result.


Q: How long can one sampler render be?

A: One window is about 12 seconds, with supported frame counts from 90 to 345 at 24 fps. Longer videos need to be split into windows; the MCP workflow can do this.


Q: What hardware does the reference sampler require?

A: The README specifies the MiniMax-H3 base and text encoder, bf16, and one GPU with at least 80 GB of memory. It points to ComfyUI assets for smaller GPUs but does not specify their performance.


Q: Does the model check whether a person consented to appear?

A: No. Identity comes from the painted reference frames, and the model cannot verify identity or consent. The README advises using likenesses only with consent and labeling generated content as AI-generated.


Q: What license applies to the weights?

A: The weights, including the int8 and GGUF files, are derivatives of MiniMax-H3 and are covered by the MiniMax H3 Community License. The supplied inference, example, ComfyUI, and MCP code is Apache 2.0.


This is a simplified guide to an AI model called Viggle-Remake maintained by Viggle.