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。
44 篇论文 · 按点赞排序
Jiawei Ma, Po-Yao Huang, Saining Xie +5 authors
Mixture of Data Experts (MoDE) enhances CLIP's performance by clustering data into experts, reducing noise and improving zero-shot classification with lower training costs.
Weifeng Chen, Jiacheng Zhang, Jie Wu +3 authors
ID-Aligner enhances identity-preserving text-to-image generation through feedback learning, addressing challenges in identity retention, aesthetic appeal, and compatibility with parameter-efficient fine-tuning methods.
Jiteng Mu, Michaël Gharbi, Richard Zhang +4 authors
A new image representation enables spatial editing of images using diffusion models by encoding them into editable elements that produce realistic outputs when decoded.
Teng Hu, Jiangning Zhang, Ran Yi +5 authors
A novel training-free video motion transfer model, COMD, decouples camera and object motions for flexible and efficient video generation.
Chuong Huynh, Seoung Wug Oh, Abhinav Shrivastava +1 authors
MaGGIe, a new framework for human instance matting, uses transformer attention and sparse convolution to maintain computational efficiency while achieving robust performance across multiple instances.
Zhaodonghui Li, Haitao Yuan, Huiming Wang +2 authors
A new query rewriting method using large language models and contrastive learning improves query execution efficiency and robustness across various datasets.
João Monteiro, Étienne Marcotte, Pierre-André Noël +5 authors
Models using cross-attention for context conditioning in language generation outperform in-context learning and reduce memory usage compared to traditional caching methods.
Haifeng Qian, Sujan Kumar Gonugondla, Sungsoo Ha +6 authors
A system for batched speculative decoding enhances latency and throughput in generating multiple sequences from large language models, outperforming single-sequence speculative decoding in both speed and quality.
Bohao Li, Yuying Ge, Yi Chen +3 authors
A benchmark evaluating text-rich visual comprehension for Multimodal Large Language Models (MLLMs) introduces SEED-Bench-2-Plus with comprehensive human annotations across diverse scenarios.
Chenyang Zhu, Kai Li, Yue Ma +2 authors
MultiBooth enhances multi-concept image generation from text by separating the process into single-concept learning and multi-concept integration phases, using bounding boxes and cross-attention maps to improve concept fidelity and reduce inference costs.
Yalda Foroutan, Daniel Rebain, Kwang Moo Yi +1 authors
Various initialization strategies, including random initialization and structure distillation from NeRF models, can enhance Gaussian Splatting, achieving results comparable to or better than those obtained using SFM data.
Junfeng Long, Wenye Yu, Quanyi Li +3 authors
A novel adversarial training method with an $H_{\infty}$ constraint enhances quadruped robots' robustness to external disturbances, demonstrated through simulations and real-world deployments on Unitree Aliengo and A1 robots.
Jiahong Xiang, Xiaoyang Xu, Fanchu Kong +3 authors
A study reveals that LLMs, particularly those using zero-shot learning and auxiliary repair information, outperform traditional APR techniques in function-level repairs, especially when using a dual-LLM framework.
Eric Brachmann, Jamie Wynn, Shuai Chen +4 authors
A learning-based approach, ACE0, estimates camera parameters from unposed images using scene coordinate regression, rivaling feature-based SfM in accuracy.
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