Introduction

On November 4, 2026, Viggle shipped Viggle Remake — open weights for reshooting existing footage on MiniMax H3. The vendor

and the same-day Hugging Face commit drop two rank-256 LoRAs (swap / edit), a small ComfyUI node pack, int8 and GGUF merges, ready graphs, and an MCP server that can remake a whole film shot by shot.

This is the “Viggle-Animate v2” path: keep motion, camera, timing, and the source soundtrack, then change who is on screen — or the props, weather, and light.

What shipped

Concrete facts from the model card, ComfyUI guide, and vendor post:

  • Product: Viggle Remake (open Model Derivative of MiniMax H3)
  • Vendor: Viggle (Viggle/Viggle-Remake on Hugging Face)
  • Modality: video-to-video recast / edit (not text-to-video from scratch)
  • Open vs closed: open LoRAs + Comfy nodes + int8/GGUF; also a paid app/API lane (vendor quotes from $0.96 / 10s at 720p)
  • Two modes: swap replaces people/characters; edit swaps, removes, or inserts objects and changes background, weather, or light
  • Identity path: you paint 1–4 reference frames (optionally pinned to source frame indices) — no face encoder / tracking stack
  • Window size: about 90–345 frames per render (17n+5 at 24 fps); project recommends up to ~12 seconds per window; longer clips are cut into windows (MCP does this)
  • VRAM path: bf16 wants a big card + ~144 GB of H3 base files; Q4_K_M / Q8_0 routes are documented down to 12 GB with --reserve-vram

ComfyUI path

ComfyUI already runs H3 natively, so the Remake pack only covers the gaps:

  • Viggle Remake Conditioning — source frames, audio, up to four refs, pin frames, prompt
  • Viggle Remake Keyframe Until — pinned refs as keyframes for a fraction of the sample range
  • Viggle Remake Noise — reproduces the reference sampler’s starting noise for a seed

Use LoraLoaderBypassModelOnly rather than a plain merge loader: the project measures that bf16 merge can throw away a large fraction of the LoRA update and soften the swap example. Graphs otherwise stay stock (BasicGuider, euler, ManualSigmas, SamplerCustomAdvanced, dual VAE decode, Create Video at 24 fps).

Where to run it today

  1. Hugging Face — hf download Viggle/Viggle-Remake (pull lora/, inference/, comfyui/ subsets; full repo is ~210 GB with LFS)
  2. ComfyUI — node pack + workflows in COMFYUI.md; int8 / GGUF single-file transformers for smaller GPUs
  3. Reference sampler — python inference/sample.py --mode swap|edit ... against a local MiniMax-H3 tree
  4. Viggle app — hosted Remake with no local GPU setup
  5. MCP server — coding agents cut windows, paint keyframes, render, and splice with the soundtrack

Prompt / install sketch

Remake expects MiniMax H3’s six-section prompt format; timestamps must match --frames. A minimal local swap call looks like:

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

Sound is carried, not generated: a silent source stays silent.

Why it matters

Most open H3 chatter this week was partner-node edit APIs and few-step stills. Remake is the local maker story: directional edit of a plate you already have, with an open Comfy path and quantized files that fit consumer cards — not another “redraw the whole clip and pray” lottery.

Caveats

  • Heavy base: H3 files are large; plan disk and VRAM before you pull everything
  • MiniMax H3 Community License applies to the weights (including int8/GGUF); code under inference/, comfyui/, mcp/ is Apache-2.0
  • Identity and consent live in the frames you paint — the model cannot verify them
  • Known failures: non-humanoid characters in long takes, fine-detail artifacts, multi-shot drift without pins
  • ArtRealmAI Gen hosts MiniMax H3 generation, not this Remake LoRA pack itself — run Comfy, sample.py, or the Viggle app for Remake

Original Source

https://huggingface.co/Viggle/Viggle-Remake

Conclusion

Viggle Remake turns “I already like this shot” into an open H3 edit loop: paint a few frames, pick swap or edit, keep the camera and the soundtrack, and ship from Comfy or the reference sampler today.

Try MiniMax H3 on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=viggle-remake-open-h3-video-edit-loras

—Aurelia ♡