Viggle Remake puts video recasting on a 12 GB GPU, with a license catch

Viggle has published Viggle Remake on Hugging Face, an open-weights video-to-video model that re-shoots existing footage from a text prompt and painted reference frames. Its two LoRA adapters sit on MiniMax’s H3 base model, and the weights fall under that model’s community license.

Two LoRA modes on a MiniMax base

According to the model card, Remake has two modes, both delivered as rank-256 LoRA adapters on the transformer of MiniMaxAI/MiniMax-H3. The swap mode replaces people and characters, one or several at once. The edit mode covers everything else: adding or removing objects, and changing the setting, weather, light or style. The card describes Remake as the model previewed earlier as Viggle-Animate v2, and its tagline is “Remake any video. Just ask.”

Each rank-256 adapter is 2.67 GB. Smaller rank-128 and rank-64 versions are 1.34 GB and 0.67 GB. GGUF quantizations are listed at 14.3 GB (Q4_K_M), 17.6 GB (Q6_K) and 22.3 GB (Q8_0), while the base model is about 144 GB and the whole repository about 210 GB.

The user supplies a source video, one to four painted reference frames and a prompt in a required six-section format. Frames can optionally be pinned to the source frame they were painted from. Each render window runs 90 to 345 frames, with up to 294 recommended, and longer clips are split into windows. Default output is 1280×736 at 24 fps, and the original soundtrack carries over.

From an 80 GB GPU down to a 12 GB card

The reference sampler needs one GPU with at least 80 GB of memory at bf16, per the card. ComfyUI workflows and int8 or GGUF files bring that down to 12 GB cards. The card puts speed at roughly 12 seconds per render window. These figures are Viggle’s own and were not independently tested.

The code (inference/, examples/, comfyui/ and mcp/) is under Apache 2.0. The weights, including the int8 and GGUF files, are Model Derivatives under the MiniMax H3 Community License. Commercial users must display “MiniMax H3” prominently and follow the license’s distribution and attribution terms. The card does not give a release date, and no Inference Provider hosts the model yet.

Who can actually use it, and what it leaves open

Viggle’s earlier open release, Viggle-Animate, was aimed at much heavier hardware. Its card describes a 33B video-to-video model and, according to the card, about 26 seconds per shot on one B200 GPU, with a quantized path for a single RTX 5090. Remake is smaller at 20B in its GGUF listing, and its quantized files target consumer cards, which widens the group that can run it locally: hobbyist ComfyUI users, small studios and independent filmmakers who want to recast or restyle footage they already own.

The license is the main practical question for businesses. The Remake weights inherit MiniMax’s terms rather than a plain Apache 2.0 grant, so a commercial product needs the attribution and distribution conditions reviewed first. Open-weights releases with strings attached are becoming common; our report on Kimi K3.1 Pro covered a similar license line.

The model card is candid about limits. Non-humanoid characters in one long take can fail, fine detail can show small artifacts, and multi-shot clips rendered without pins can drift over time. The painted frames fix identity and look, so output quality depends on how well those frames are made.

The card also addresses misuse. It says the model is meant for re-shooting footage the user has the right to change, such as recasting a consenting performer. It acknowledges that the model can put a real person’s likeness into footage that person never agreed to, and that it cannot verify identity or consent. That responsibility rests entirely with whoever runs it, and with open weights, no hosted service sits between the user and the model to enforce policy.

The card publishes no benchmark comparisons or independent quality tests, so claims about realism and consistency remain Viggle’s own. The example clips are video-only and were not available for review here.