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。
40 篇论文 · 按点赞排序
Suttisak Wizadwongsa, Worameth Chinchuthakun, Pramook Khungurn +2 authors
Proposed Heavy Ball and Generalized Heavy Ball momentum techniques improve numerical stability and reduce artifacts in diffusion model sampling, enhancing image quality.
Zihan Zhang, Richard Liu, Kfir Aberman +1 authors
The work adapts DDPM for motion synthesis by extending the diffusion framework to a temporal axis, enabling the generation of long streams of clean motion frames.
Adrien Bardes, Jean Ponce, Yann LeCun
MC-JEPA, a joint-embedding predictive architecture, simultaneously learns optical flow and content features through self-supervised learning, improving both optical flow estimation and downstream tasks like semantic segmentation.
Dehua Zheng, Wenhui Dong, Hailin Hu +2 authors
Focus-DETR enhances DETR-like models' performance and efficiency by implementing dual attention that focuses on informative tokens with localization and semantic information.
Quankai Gao, Qiangeng Xu, Hao Su +2 authors
Strivec models a 3D scene as a radiance field using sparsely distributed and compactly factorized local tensor grids via CP decomposition, outperforming methods like TensoRF and Instant-NGP with fewer parameters.
Reuben Tan, Matthias De Lange, Michael Iuzzolino +4 authors
Multiscale Video Pretraining (MVP) enhances long-term activity forecasting by learning robust representations across different timescales, improving accuracy on tasks like action anticipation and video summary prediction.
Richard Yuanzhe Pang, Stephen Roller, Kyunghyun Cho +2 authors
Improvements in social conversational agents are achieved by learning from user interactions without annotations, using signals like response length, sentiment, and reactions to measure utterance quality.
Zhipeng Huang, Zhizheng Zhang, Cuiling Lan +3 authors
A new Adaptive Frequency Filtering (AFF) token mixer based on Fourier transforms improves accuracy and efficiency in lightweight neural networks for visual tasks.
Kyle Hatch, Benjamin Eysenbach, Rafael Rafailov +4 authors
A method for offline, example-based control learns an implicit model of multi-step transitions, outperforming reward function-based methods and demonstrating robustness and scaling with dataset size.
Boris Ivanovic, Guanyu Song, Igor Gilitschenski +1 authors
trajdata provides a unified interface for trajectory and map data, facilitating research across multiple human trajectory datasets for autonomous vehicles and pedestrian motion tracking.
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