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。
45 篇论文 · 按点赞排序
Mamshad Nayeem Rizve, Fan Fei, Jayakrishnan Unnikrishnan +5 authors
VidLA, a simple yet effective video-language alignment approach, uses temporally hierarchical data tokens and a two-tower architecture to enhance performance with large, semantically aligned datasets and surpass state-of-the-art methods.
Badri N. Patro, Vijay S. Agneeswaran
SiMBA, a new state-space model architecture using Einstein FFT for channel modeling and Mamba blocks for sequence modeling, outperforms existing SSMs and transformers across image and time-series benchmarks.
Qiao Gu, Zhaoyang Lv, Duncan Frost +3 authors
EgoLifter segments 3D scenes captured from egocentric sensors into individual objects using 3D Gaussians, SAM for weak supervision, and a transient prediction module for dynamic object filtering.
Huawei Wei, Zejun Yang, Zhisheng Wang
AniPortrait generates high-quality portrait animations from audio and a reference image using 3D intermediate representations, 2D facial landmarks, and a diffusion model with a motion module.
Trong-Tung Nguyen, Duc-Anh Nguyen, Anh Tran +1 authors
FlexEdit is a flexible and controllable object-centric editing framework that uses iterative latent adjustment and adaptive masking to improve realism and control in object replacement and insertion.
Yuanze Lin, Ronald Clark, Philip Torr
DreamPolisher uses Gaussian Splatting with geometric guidance to generate detailed and consistent 3D objects from text descriptions, improving upon existing text-to-3D methods.
Ruining Li, Chuanxia Zheng, Christian Rupprecht +1 authors
DragAPart generates new images based on part-level interactions using fine-tuned image generators and a new synthetic dataset, showing improved part-level motion understanding compared to previous methods.
Orion Weller, Benjamin Chang, Sean MacAvaney +5 authors
FollowIR dataset and evaluation framework demonstrate that large language models can be fine-tuned to better follow detailed instructions in information retrieval, improving performance over baseline models.
Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan +5 authors
StreamingT2V is an autoregressive text-to-video diffusion model that generates long, high-quality videos with smooth transitions by incorporating short-term and long-term memory blocks and a randomized blending approach.
Kai Yuan, Christoph Bauinger, Xiangyi Zhang +6 authors
A SYCL implementation of MLPs optimized for Intel's Data Center GPU demonstrates superior performance compared to CUDA across various applications, including image compression, neural radiance fields, and physics-informed machine learning.
Chaoyun Zhang, Zicheng Ma, Yuhao Wu +12 authors
Allhands is an innovative feedback analysis framework using large language models for classification, topic modeling, and query interpretation to deliver comprehensive insights via a natural language interface.
Kevin Xie, Jonathan Lorraine, Tianshi Cao +5 authors
LATTE3D generates high-quality 3D objects quickly by using a scalable architecture and 3D-aware diffusion priors.
Yang Chen, Yingwei Pan, Haibo Yang +2 authors
A novel Visual Prompt-guided text-to-3D diffusion model (VP3D) boosts text-to-3D generation by integrating visual appearance knowledge from 2D images.
Dejan Grubisic, Chris Cummins, Volker Seeker +1 authors
A feedback-driven Large Language Model optimizes LLVM IR for code size, improving upon existing methods with iterative refinement and sampling techniques.
Rricha Jalota, Lyan Verwimp, Markus Nussbaum-Thom +3 authors
Adapter bottlenecks are used to integrate dialect-specific characteristics into a World English NNLM for virtual assistants, improving accuracy while maintaining low latency and memory usage.
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