Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
47 篇论文 · 按点赞排序
Zeyu Zhang, Akide Liu, Ian Reid +3 authors
Motion Mamba, utilizing a hierarchical and bidirectional SSM-based architecture, enhances motion generation by improving quality and speed on HumanML3D and KIT-ML datasets.
Wenhao Wang, Yi Yang
A large-scale dataset of text-to-video prompts and generated videos opens new research avenues for improving text-to-video diffusion models.
Yufei Zhan, Yousong Zhu, Hongyin Zhao +3 authors
Griffon v2, a unified high-resolution model with a down-sampling projector and visual tokenizer, achieves state-of-the-art performance in object detection, counting, and referring tasks by overcoming image resolution and multimodal perception limitations.
Guo Chen, Yifei Huang, Jilan Xu +7 authors
The Video Mamba Suite demonstrates the strong potential of Mamba, a state space model, in video understanding tasks, showing its efficiency and performance advantages over Transformers.
Samir Yitzhak Gadre, Georgios Smyrnis, Vaishaal Shankar +20 authors
The study investigates scaling laws in over-trained language models and relates model perplexity to downstream task performance, providing predictions with reduced computational cost.
Yue Ma, Yingqing He, Hongfa Wang +8 authors
A framework named Follow-Your-Click allows precise image-to-video animation through user clicks and short motion prompts, enhancing controllability and local animation quality.
Jonathan Heek, Emiel Hoogeboom, Tim Salimans
Multistep Consistency Models unify Consistency Models and TRACT, offering a trade-off between sampling speed and quality, achieving high sample quality in fewer steps compared to conventional methods.
Nan Jiang, Zhiyuan Zhang, Hongjie Li +6 authors
A new diffusion-based autoregressive model using the TRUMANS dataset efficiently generates human-scene interaction motion sequences with high realism and zero-shot generalizability.
Wejia Wu, Zhuang Li, Yuchao Gu +7 authors
DragAnything uses entity representation for motion control in video generation, offering user-friendly interaction and distinct motion control for multiple objects.
Haoyu Zhen, Xiaowen Qiu, Peihao Chen +5 authors
3D-VLA integrates 3D perception, reasoning, and action through a generative world model, improving capabilities in embodied environments with the help of embodied diffusion models.
Keith Rush, Zachary Charles, Zachary Garrett
FAX is a JAX-based library enabling distributed and federated computations using TPUs, XLA HLO, and native federated automatic differentiation.
Xinjie Zhang, Xingtong Ge, Tongda Xu +6 authors
Gaussian Splatting, a novel image representation and compression method, achieves high visual quality and fast rendering speeds with lower GPU memory usage and faster fitting time compared to existing INRs.
Rajat Chawla, Arkajit Datta, Tushar Verma +6 authors
Veagle, a novel multimodal model, enhances the integration of visual and textual information by dynamically projecting visual data into language models, outperforming existing models in visual question answering and image understanding tasks.
Chris Kelly, Luhui Hu, Jiayin Hu +7 authors
The VisionGPT-3D framework integrates advanced computer vision and large language models to automatically generate optimal 3D visuals from 2D images and text prompts.
Tao Huang, Xiaohuan Pei, Shan You +3 authors
A novel local scanning strategy for Vision Mamba improves vision task performance by optimizing scan directions and maintaining local 2D dependencies across distinct layers.
Ri-Zhao Qiu, Yafei Hu, Ge Yang +8 authors
GeFF, a generalizable neural feature field, unifies navigation and manipulation tasks by leveraging generative novel view synthesis and CLIP feature distillation for real-time performance in dynamic scenes.
Pengchong Qiao, Lei Shang, Chang Liu +3 authors
A plug-and-play method called Subject-Derived regularization (SuDe) ensures that subject-driven text-to-image generation inherits category attributes while learning subject-specific ones, enhancing attribute-related generation quality.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号