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
Kan Zhu, Yilong Zhao, Liangyu Zhao +12 authors
NanoFlow enhances serving throughput by exploiting intra-device parallelism through operation co-scheduling and execution unit scheduling on GPUs.
Peng Xing, Haofan Wang, Yanpeng Sun +5 authors
A large-scale style transfer dataset and end-to-end training model are presented that effectively decouple content and style features for enhanced image generation control.
Keqiang Sun, Amin Jourabloo, Riddhish Bhalodia +9 authors
A text-conditioned generative model creates photo-realistic 3D avatars with complete facial details and controllable expressions using latent diffusion models and a prior expression driving model.
Jainaveen Sundaram, Ravishankar Iyer
LLaVaOLMoBitnet1B is the first open-source ternary multimodal LLM that processes images and text to generate coherent responses, designed to run efficiently on small computational resources.
Shengyuan Zhang, Ling Yang, Zejian Li +6 authors
DisBack, an extended score distillation method, accelerates diffusion model sampling by leveraging the entire convergence trajectory of the teacher model, improving student generator performance and reducing the score mismatch issue.
Peng Wang, Zhaohai Li, Jun Tang +4 authors
Platypus is a unified text reading model that combines the versatility of generalist models with the high accuracy and efficiency of specialist models, demonstrated on various benchmarks.
Fuwen Tan, Royson Lee, Łukasz Dudziak +5 authors
MobileQuant facilitates the on-device deployment of large language models using integer-only quantization, reducing latency and energy consumption while maintaining high accuracy.
Hengyi Wang, Lourdes Agapito
Spann3R is a transformer-based method for dense 3D reconstruction from image collections, using an external spatial memory to predict global pointmaps and demonstrating real-time performance and generalization.
Lean Wang, Huazuo Gao, Chenggang Zhao +2 authors
Loss-Free Balancing improves performance and load balance in Mixture-of-Experts models by dynamically adjusting expert biases without introducing auxiliary loss gradients.
Rui Zhang, Yaosen Chen, Yuegen Liu +3 authors
A novel training-free method for generating smooth transition videos using video-level diffusion models, Gaussian Process Regression, and a Frequency-aware Bidirectional Fusion architecture.
Kaizhao Liang, Bo Liu, Lizhang Chen +1 authors
A new family of memory-efficient optimizers called Online Subspace Descent provides convergence guarantees and outperforms existing methods for pretraining large language models.
Yi Liu, Junzhe Yu, Huijia Sun +4 authors
ToxicDetector is a lightweight greybox method using LLMs, embeddings, and MLP classifiers to efficiently detect toxic prompts with high accuracy and low false positives.
Wenshuo Peng, Kaipeng Zhang, Sai Qian Zhang
A novel text-guided 3D motion synthesis method called T3M provides precise control over motion generation, offering improved quality and diversity compared to existing speech-driven approaches.
Jillian Fisher, Skyler Hallinan, Ximing Lu +3 authors
StyleRemix is an adaptive and interpretable authorship obfuscation method using pre-trained LoRA modules to perturb specific stylistic elements, outperforming state-of-the-art methods and larger LLMs.
Uri Katz, Mosh Levy, Yoav Goldberg
Knowledge Navigator uses large language models and clustering methods to organize and navigate scientific literature efficiently.
Danlong Yuan, Jiahao Liu, Bei Li +4 authors
ReMamba addresses Mamba's long-context efficiency limitations through selective compression and adaptation, narrowing the performance gap with transformer models on long contexts.
Tao Wu, Yong Zhang, Xintao Wang +5 authors
CustomCrafter enhances video diffusion models to preserve motion generation and concept combination without additional fine-tuning or guiding videos.
Jesse C. Cresswell, Taewoo Kim
Improved implementation and algorithmic enhancements of XGBoost in diffusion and flow-matching models enable scaling to larger datasets and superior performance on scientific benchmarks.
Letian Fu, Huang Huang, Gaurav Datta +5 authors
In-Context Robot Transformer (ICRT) enables flexible and training-free execution of new tasks by interpreting sensorimotor trajectories during input phase without updating policy parameters.
Xu He, Xiaoyu Li, Di Kang +7 authors
MagicMan, a multi-view diffusion model using SMPL-X and hybrid multi-view attention, generates consistent high-quality novel views and enhances 3D reconstruction accuracy with an iterative refinement strategy.
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