Character-replacement video pipelines have accumulated a lot of machinery. A pose estimator to get the skeleton. A segmentation model for masks. Sometimes a face tracker. A text encoder to hold a prompt nobody is confident about. Each stage is another model to load, another failure mode, another thing that drifts when the subject turns.
Viggle-Animate, released September 4, 2026 as Viggle Research’s first open-weight model, deletes all of it. The interface is two files: a driving video, and one frame from that same video with the character repainted in any image editor. The model propagates that edit across the shot.
Why the repainted frame works
The insight is that a frame from the source footage already encodes everything the auxiliary models were being asked to reconstruct — the pose, the camera, the lighting, the occlusions, the frame’s own geometry. Repaint the character in place and you’ve handed the model a target appearance that is already registered to the scene. There’s nothing left to estimate.
So appearance comes from the paint, geometry and motion come from the driving video, and the text encoder isn’t loaded at all — the model runs on a frozen text embedding.
The architecture
Viggle-Animate is a full-parameter finetune of MiniMax-H3’s ref2va transformer, 33.1B parameters, jointly distilled with Distribution Matching Distillation down to three forward passes. The distillation uses a dual-teacher setup, different teachers at high and low noise levels, trading off replacement fidelity against overall visual quality.
The speed numbers follow from the step count: 124 frames at 480×832, 24 fps, in 26 seconds on a single B200, which Viggle measures at 6.1× faster per clip than Wan2.2-Animate-14B — three forward passes against 40+, and zero auxiliary models against several.
The team reports its biggest quality margins in exactly the cases these pipelines usually break: fast motion and extreme poses, where a pose estimator’s errors compound. The same interface also handles animals, stylized characters, and some non-humanoid objects, since nothing in it assumes a human skeleton.
Getting it running
Weights are on Hugging Face under the MiniMax H3 Community License (the weights are a Model Derivative); the inference/ and examples/ code is Apache 2.0. Hardware is the wall for most people: 96 GB of VRAM at bf16. Quantized to NVFP4 with compilation, it will run on a 32 GB RTX 5090. There’s a hosted Space for anyone who’d rather just try it.
ComfyUI users got there quickly — community ports appeared within days, and our Hugging Face tracker has been logging steep download growth on the ComfyUI-packaged conversions all week.
What it can’t do
Viggle is direct about the limits. The model inherits your edit exactly: it cannot invent detail the paint doesn’t show, so a sloppy repaint stays sloppy for the whole clip. Lip-sync is weak in close-ups — mouth shapes don’t track speech tightly, though there’s an --audio conditioning flag. Multiple characters, close physical interaction, and shot changes all degrade quality. And identity can drift when a subject leaves frame or moves far from the reference pose, settling back toward the driving video’s original appearance.
The license also carries obligations worth reading before anyone ships work with it: consent from depicted individuals, and labeling outputs as AI-generated.