PhyMotion accepted at NeurIPS 2026: 3D physics rewards lift 1.3B video model to 14B level
mohitban47 · x · 2026-09-29
PhyMotion, a structured 3D motion reward for physics-grounded human video generation, is accepted at NeurIPS 2026. It lifts generated videos into 3D and scores motion via physics simulation on kinematic plausibility, contact/balance consistency, and dynamic feasibility — where 2D rewards misleadingly reward floating feet and limb penetration. RL post-training with PhyMotion brings a 1.3B model to 14B-level human preference. Updated paper, code, and demos coming.
More from Multimodal
- Inception Labs launches Mercury Voice, a diffusion LLM with 2x lower latency than GPT-6 Luna and Claude Haiku 4.5 — StefanoErmon · 2026-09-30
- ComfyUI-Continuity Update Adds Cast Management, RefMod Presets and Seam Optimization — Okims_kor · 2026-09-30
- ComfyUI-Continuity Update Adds Cast Management, RefMod Presets and Seam Optimization — Okims_kor · 2026-09-30
- AI turns its subagents into a rock band, full MV made with local music and video models — WolframRvnwlf · 2026-09-30
- Building a Codex-to-Flux 3 Pipeline That Autonomously Generates Ancient Greece Memes — venturetwins · 2026-09-30
- Humans Only: DIGI, an AI Singer Who Built Our World and Can't Get In (Anime MV) — wanderingtofu · 2026-09-30