Fatih Porikli

Generating videos that simultaneously obey multiple interacting physical laws, like buoyancy and fluid dynamics, is now achievable with a novel RL framework that explicitly aligns temporal dynamics and global scene coherence.

High-quality mobile video generation is now possible with a 5B-parameter model, challenging the notion that smaller models are necessary for mobile applications.

CARVE's chunk-adaptive reranking strategy reveals that tailored retrieval configurations can dramatically improve video generation performance.

Diffusion language models have surprisingly redundant early layers, enabling nearly 20% FLOPs reduction at inference time via layer skipping without sacrificing performance.