논문리뷰 mllms mllm Training-Free 논문 리뷰 CVPR diffusion codebook vq-vae paper review 3D Vision video mllm fine-grained video motion understanding motionvid motionsight token compression token pruning GlobalCom Zoom-Refine T2M SMLP-X Human estimation FreeVA text-to-motion 3D pose SOKE Auto-regressive Diffusion AR-Diffusion paper-review visual tokenizer va-gan textdiffusion ecp vqct pose generation unipose motion generation vision encoder iclr motion estimation DeepLearning ISHIHARA 딥러닝 SLP Multimodal Rewind 경량화 salad egocentric Cosmos nlp DataSet Real-time Google