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。
50 篇论文 · 按点赞排序
Sangwon Jang, June Suk Choi, Jaehyeong Jo +2 authors
The Silent Branding Attack manipulates text-to-image diffusion models to generate images with specific logos without text triggers, using data poisoning techniques.
Zeyu Zhang, Yiran Wang, Wei Mao +7 authors
Motion Anything improves multimodal motion generation through attention-based masking and adaptive condition encoding, achieving state-of-the-art results across benchmarks.
Wanhua Li, Renping Zhou, Jiawei Zhou +5 authors
A new method, 4D LangSplat, uses multimodal object-wise video captions and a status deformable network to generate precise and efficient 4D language fields for dynamic scenes.
Jongwoo Ko, Tianyi Chen, Sungnyun Kim +4 authors
A contrastive distillation method (DistiLLM-2) enhances student language models by optimizing teacher and student responses differently, improving performance across various tasks and applications.
Shitong Shao, Zikai Zhou, Dian Xie +4 authors
CoRe^2, a new plug-and-play inference paradigm, enhances both the efficiency and quality of text-to-image generative models by leveraging classifier-free guidance and weak-to-strong refinement.
James Burgess, Xiaohan Wang, Yuhui Zhang +5 authors
VidDiff is a new task for identifying subtle differences between videos of the same action, addressed by a benchmark dataset and a three-stage method utilizing specialized foundation models.
Hyeonho Jeong, Suhyeon Lee, Jong Chul Ye
Reangle-A-Video generates synchronized multi-view videos from a single input video using image-to-video diffusion transformers and cross-view consistency guidance.
Sangwoo Park, Seanie Lee, Byungjoo Kim +1 authors
FedRand framework enhances data privacy in federated learning by keeping a subset of LoRA parameters private, reducing the risk of membership inference attacks while maintaining model accuracy.
Zhixuan Lin, Evgenii Nikishin, Xu Owen He +1 authors
A Forgetting Transformer (FoX) improves performance on long-context language modeling and short-context tasks by incorporating a forget gate mechanism into attention scores, while maintaining compatibility with FlashAttention and superior long-context handling.
Junzhe Li, Xuerui Qiu, Linrui Xu +5 authors
UniF<sup>2</sup>ace is a unified multimodal model that excels in fine-grained face understanding and generation using a two-level mixture-of-experts and diffusion techniques.
Wenxuan Huang, Bohan Jia, Zijie Zhai +5 authors
A reasoning MLLM called Vision-R1 is proposed to enhance multimodal reasoning capabilities using Reinforcement Learning and a constructed multimodal CoT dataset, achieving high performance on multimodal math reasoning benchmarks.
Liang Wen, Yunke Cai, Fenrui Xiao +11 authors
The Light-R1 series develops strong long-chain-of-thought models through curriculum training, semi-on-policy DPO, and reinforcement learning, achieving state-of-the-art performance in math and general domains.
Yuxuan Zhang, Yirui Yuan, Yiren Song +2 authors
EasyControl enhances diffusion transformers with efficient and flexible condition handling, position awareness, and causal attention for high-performance image synthesis.
Yongsheng Yu, Ziyun Zeng, Haitian Zheng +1 authors
OmniPaint, utilizing a unified framework with a diffusion prior and CycleFlow, achieves high-fidelity object removal and insertion while maintaining scene geometry and object properties.
Yuxuan Bian, Zhaoyang Zhang, Xuan Ju +4 authors
VideoPainter uses a dual-stream paradigm with a context encoder to enhance video inpainting by balancing background context and foreground generation, supporting any-length inpainting and various downstream applications.
Muzhi Zhu, Yuzhuo Tian, Hao Chen +5 authors
HLMAT introduces a new paradigm for assessing fine-grained pixel understanding in MLLMs through interactive segmentation tasks, enabling high-quality mask generation and supporting additional tasks without architectural changes.
Subin Kim, Seoung Wug Oh, Jui-Hsien Wang +2 authors
SynCoS, a novel inference framework for text-to-video diffusion models, enhances long video generation by synchronizing denoising paths for smooth transitions and global coherence.
Giulio Corallo, Orion Weller, Fabio Petroni +1 authors
Task-aware key-value cache compression for large language models enhances reasoning efficiency and accuracy over retrieval-augmented generation and task-agnostic compression methods.
Huatong Song, Jinhao Jiang, Yingqian Min +5 authors
R1-Searcher is a reinforcement learning approach that enhances large language models' reasoning by autonomously accessing external knowledge, achieving better performance than existing methods.
Sinuo Liu, Chenyang Lyu, Minghao Wu +3 authors
Large Reasoning Models leveraging Chain-of-Thought reasoning have redefined Machine Translation by transforming it into a dynamic reasoning task that emphasizes contextual coherence, cultural intent, and self-reflection, demonstrating significant improvements in various translation scenarios.
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