Viggle-Animate open-sourced: replace video characters from one repainted frame, no pose estimator
cocktailpeanut · x · 2026-09-07
Viggle released Viggle-Animate on Hugging Face: a 33.1B full finetune of MiniMax-H3's ref2va transformer, distilled with DMD to just three forward passes.
- Workflow: repaint the character into one frame of the video; the model propagates the edit across the shot, keeping motion, camera and timing untouched
- No pose estimator, segmentation mask, face tracker or text encoder needed — two inputs, three forward passes, 26 seconds per shot on one GPU
- Since the reference is a frame of the clip itself, pose, framing and lighting align naturally; the model never needs to be told what the new character is
- Strong where replacement is hardest: fast motion like head whips, full kicks and jumps, tracked frame by frame
- cocktailpeanut's hands-on test found quality exceeded expectations — three steps on a local PC
Released under the minimax-h3-community-license.
More from Multimodal
- MiniMax H3 Turbo Cinematic Renders an Impressive Anime Fight Scene — Ok-Vegetable-2455 · 2026-09-07
- Minimax H3 fight-scene test: fluid martial arts motion but unstable Sharingan details — Ok-Vegetable-2455 · 2026-09-07
- AI-generated Spanish original song 'Una vida en tu aroma' released in full — azed_ai · 2026-09-07
- 62-second AI short built from 14 Grok Imagine clips with a full post pipeline — tetsuoai · 2026-09-07
- Tech launch videos overuse Suno lofi; time to commission real musicians — eschadiol · 2026-09-07
- GPT-6 Astra builds interactive 3D Seoul with 267,000 buildings in 43 minutes — DeryaTR_ · 2026-09-07