StrideWorld Benchmark Finds Most Interactive Video World Models Deform Scenes
SJTU · hf · 2026-11-19
Researchers from SJTU and collaborators released StrideWorld, a benchmark that fixes references before motion to detect scene deformation and action inconsistency in interactive video world models. Results show most models deform scenes, and only depth-scaled command models preserve composed routes.
More from Research
- Uzbek-language corpus trends on Hugging Face: 1M-10M entries, ODC-BY licensed — Digitaluz · 2026-11-19
- WaveView Injects Warped Views Into Noise Phase, Beating NVS Baselines at 2.1x Speed — KIT-Karlsruhe · 2026-11-19
- ICLR 2027 rebuttal tracker: only 39.4% of papers changed score, 25:1 up vs down — snu95 · 2026-11-19
- Chris Olah's decade of persistence: from DeepDream to founding Distill and interpretability — arvindsatya1 · 2026-11-19
- Ropedia Xperience 2: egocentric video plus metric 3D reconstructions hits HF trending for Physical AI — liuziwei7 · 2026-11-19
- Four slope samples, weights 1-2-2-1: a visual explainer of why Runge-Kutta 4 is so accurate — burny_tech · 2026-11-19