Viggle Remake

Remake any video. Just ask.

👉 Try Viggle Remake in the Viggle app: no setup, no GPU

_{La La Land. One shot from La La Land (2016), remade 22 ways: recast, new outfits, weather, things added and removed,
a new style. Unmute to hear it. Footage and music from La La Land (2016), © its owners. Not affiliated with or endorsed by them.}

_{White House Funk. A 116-second music video recast with lora/swap, 69 cuts, original
soundtrack (unmute). AI-generated parody: the people shown did not take part in, approve or endorse it.
Source: Mark Ronson ft. Bruno Mars, Uptown Funk (2014). To make one yourself, see the MCP.}

_{Burly Brawl. The Matrix Reloaded (2003), Neo as a tuxedo cat and every Agent Smith as a
shiba inu, 101 seconds with lora/swap and the original soundtrack.}

Viggle Remake is built agent-first. Give an agent (Claude Code, codex, anything that speaks MCP) a video and one sentence: "recast the band as US politicians", "make it snow", "turn everyone into cats". It finds the cuts, paints the keyframes, renders every window, checks its own work and hands back the finished film, with the original motion, camera, timing and soundtrack. Set up the agent.

Recast people · swap outfits · add or remove objects · change the weather, light, setting or style · whole films, not just clips.

Prefer to drive it by hand? Write a prompt, paint one to four frames, and sample.py renders the shot.

Remake is the "Viggle-Animate v2" previewed on the Viggle-Animate card: Animate replaces a character; Remake also swaps, removes and inserts objects, and changes the setting, the weather or the light.

Quick start

Just want to use it? The Viggle app runs Remake for you. The rest of this section is for running it on your own GPUs.

You need the base model (about 144 GB of its files) and this repository's LoRAs and sampler:

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"   # 5.5 GB; the whole repo is ~210 GB
pip install -r Viggle-Remake/requirements.txt   # plus an ffmpeg binary on PATH

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

sample.py renders one window of 90–345 frames (--frames = 17n + 5, at 24 fps; we recommend up to 294) with the source soundtrack; --short-edge 736 renders 1280×736. Its defaults are the settings every clip here used. examples/demo.sh reruns one example of each mode from examples/inputs/: if the turbines are still there, or the pizza maker is unchanged, the LoRA did not load.

Smaller LoRAs: --rank 128 or --rank 64 loads lora_lowrank/ (1.34 / 0.67 GB), near-identical output.

ComfyUI and smaller GPUs

Node pack, workflows, and int8/GGUF files down to 12 GB: COMFYUI.md.

Agent: the MCP server

mcp/ lets a coding agent (Claude Code, codex) remake a whole film: it cuts the windows, paints the keyframes, writes the prompts, renders on your GPUs and splices the result with the soundtrack.

hf download Viggle/Viggle-Remake --include "mcp/*" --local-dir viggle-remake   # the MCP only, not the weights
cd viggle-remake/mcp
pip install -e .              # the server
pip install -e '.[worker]'    # GPU machines: also the render worker

Do not git clone this repository: with git-lfs it downloads all ~210 GB of weights.

To remake White House Funk (source, eight characters and task in mcp/examples/white-house-funk/):

for g in 0 1; do viggle-remake-worker --gpu $g > worker_$g.log 2>&1 & done   # one render worker per GPU
examples/white-house-funk/run.sh     # codex (ChatGPT plan login) runs the whole film: ~1.5 h on two GPUs

The agent paints its own keyframes, so every run differs in detail. Setup, tools and paint backends: mcp/README.md.

Two modes

Each mode is a LoRA (rank 256) on MiniMax-H3's base transformer; --mode picks one.

--mode swap (lora/swap): replace people and characters, one or several at once.
--mode edit (lora/edit): everything else: swap, remove or insert objects, change the background, weather or light.

Inputs: painted frames and the prompt

Pass one to four painted frames (--ref, called <Picture N> in the prompt), optionally each pinned to the source frame it was painted from (--pin), and a prompt in MiniMax-H3's six-section format. How to paint them, when to pin, and how to write the prompt: PROMPTING.md.

Limitations

About 12 seconds per render. Longer videos are split into windows, at cuts where possible (the MCP does this).
Heavy. The 144 GB base and its text encoder are required. sample.py needs one 80 GB+ GPU at bf16; see COMFYUI.md for smaller cards.
The six-section prompt is required (PROMPTING.md); timestamps must fit --frames.
The painted frames decide identity and look; pinned frames also fix the composition at that frame.
Sound is carried, not made: the output has the source soundtrack; a silent source gives a silent result.
Known failures: a non-humanoid character in one long take; small artifacts in fine detail; shots drifting in time when a multi-shot clip is rendered without pins.

Intended use

This model exists to re-shoot footage you have the right to change: recasting a consenting performer, or changing a set, a prop or the weather. It will just as readily put a real person's likeness into footage they never agreed to appear in. Note where that decision is made: the identity comes from the frames you paint. Any image editor's safeguards sit upstream of this model, and none of them are in it. It cannot verify identity or consent. Do not run it on people who have not agreed to it, label what you generate as AI-generated, and see Section V.5 of the Agreement if you offer this as a service.

Example footage

The source footage in examples/ and mcp/examples/ is third-party, © its owners, and credited in examples/CREDITS.md.

What this repository contains

path	contents
`lora/swap/`, `lora/edit/`	the two modes, PEFT LoRA, rank 256 (2.67 GB each)
`lora_lowrank/`	the same at rank 128 and 64
`comfyui/`, `int8/`, `gguf/`	ComfyUI LoRAs, node pack and workflows; merged int8 and GGUF transformers. See `COMFYUI.md`
`inference/`, `requirements.txt`	`sample.py`, the reference sampler, and its package set
`examples/`	`demo.sh`, the example prompts and inputs, the clips on this card
`mcp/`	the MCP server and render worker, with a full example run

License

The weights, including the merged int8/ and gguf/ files, are a Model Derivative of MiniMax H3, so the MiniMax H3 Community License applies to them. Read it before you redistribute them or ship a product on them. What we added is listed in MODIFICATIONS.md.

The code in inference/, examples/, comfyui/ and mcp/ is Apache 2.0 (LICENSE-CODE). That licence covers the scripts and prompts only. Neither licence grants any rights to the third-party footage in examples/media/, examples/inputs/ and mcp/examples/, or to the likenesses of the people shown.

Attribution

If you build a product or service on Viggle Remake, please credit it where your users will see it, for example in the UI or the credits:

Powered by Viggle Remake (https://huggingface.co/Viggle/Viggle-Remake)

Separately, if you ship a product or service on these weights, the MiniMax H3 Community License requires you to:

display "MiniMax H3" prominently in the user interface of a commercial product or service (Section IV.2);
give a copy of the License to everyone you distribute the weights to, mark any files you modified, and include the NOTICE file (Section III);
get written authorization from MiniMax first if your products and services make more than US$20 million a year (Section IV.1).

Citation

@misc{viggle2026remake,
  title  = {Viggle Remake: Remake Any Video. Just Ask.},
  author = {Viggle Research},
  year   = {2026},
  url    = {https://huggingface.co/Viggle/Viggle-Remake}
}

Model tree for Viggle/Viggle-Remake

Base model

MiniMaxAI/MiniMax-H3

Adapter

(188)

this model

Viggle
/

Viggle-Remake