ICML oral Motive: first motion attribution framework for video generation, 74.1% win rate
cindy_x_wu · x · 2026-09-04
The ICML 2026 oral 'Motion Attribution for Video Generation' (Xindi Wu, Antonio Torralba, Sanja Fidler, Jonathan Lorraine, et al.) presents Motive, the first gradient-based framework attributing motion (rather than visual appearance) in video generation models. Using motion-weighted loss masks to isolate temporal dynamics, Motive identifies which fine-tuning clips improve or degrade motion and guides data curation, improving motion smoothness and dynamic degree on VBench with a 74.1% human preference win rate over the base model.
Related event: ICML oral paper on motion attribution for video generation released(2 posts)→
More from Multimodal
- Makepad flow: a Rust single-executable ComfyUI alternative built in one day — anselm · 2026-09-04
- RTX 4060 Ti 16GB runs MiniMax video model locally: 768p in 3 minutes — aziib · 2026-09-04
- LTX 2.5 gains traction for AI music videos, runs smoothly on Macs — cocktailpeanut · 2026-09-04
- 'Sometimes I wish I live my life as AI': MiniMax H3 pixel-art girl goes viral — aziz4ai · 2026-09-04
- MiniMax H3 generates lifelike pixel-game-style video, upscaled with Magnific — MarioKrenn6240 · 2026-09-04
- Dev rebuilds Vine with H3 Max Turbo on fal: an infinite video feed that generates faster than you scroll — chrisfirst · 2026-09-04