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
Xiaozheng Zheng, Chao Wen, Zhaohu Li +9 authors
A framework for creating 3D photorealistic and animatable head avatars using prior learning and few-shot personalization with Gaussian Splatting and dynamic modeling.
Junjie He, Yifeng Geng, Liefeng Bo
UniPortrait is a framework for high-fidelity human image personalization using diffusion models, featuring both single- and multi-ID customization and diverse layout options.
Qizhen Zhang, Nikolas Gritsch, Dwaraknath Gnaneshwar +8 authors
BAM enhances the Mixture of Experts framework by fully recycling dense model parameters, including attention layers, leading to better performance and efficiency than previous methods.
Junxian Li, Di Zhang, Xunzhi Wang +14 authors
ChemVLM, a multimodal language model combiningtext and image understanding in chemistry, achieves state-of-the-art performance using VIT-MLP-LLM, ChemLLM-20B, and InternVIT-6B architectures.
Guan Luo, Tian-Xing Xu, Ying-Tian Liu +3 authors
A novel single-image-driven 3D scene editing approach using 3D Gaussian Splatting achieves intuitive manipulation via direct 2D image editing, capturing long-range and non-rigid deformations with superior quality.
Sang-Hoon Lee, Ha-Yeong Choi, Seong-Whan Lee
PeriodWave-Turbo, an enhanced adversarial flow matching model with a pre-trained CFM generator, improves waveform generation efficiency and high-frequency detail, achieving state-of-the-art performance with reduced ODE steps.
Shachar Don-Yehiya, Leshem Choshen, Omri Abend
The ShareLM collection and plugin facilitate the voluntary sharing and rating of human-large language model conversations, enhancing open-source model development.
Nursena Koprucu, Meher Shashwat Nigam, Shicheng Xu +6 authors
Diffusion Classifier for 3D Objects (DC3DO) uses class-conditional diffusion models to achieve zero-shot classification of 3D shapes with superior multimodal reasoning.
Haoran Sun, Renren Jin, Shaoyang Xu +10 authors
FuxiTranyu, a multilingual large language model, achieves competitive performance across various benchmarks and learns consistent multilingual representations through balanced training and diverse fine-tuning.
Zhengtong Xu, Raghava Uppuluri, Xinwei Zhang +5 authors
UniT uses VQVAE for tactile representation learning, demonstrating superior performance in various tasks including pose estimation and policy learning, with zero-shot transfer capabilities.
Chenjie Cao, Chaohui Yu, Yanwei Fu +2 authors
MVInpainter redefines 3D editing as a multi-view 2D inpainting task, enhancing cross-view consistency with video priors and slot attention, enabling pose-free training and inference.
Sang-Hoon Lee, Ha-Yeong Choi, Seong-Whan Lee
PeriodWave is a novel universal waveform generation model that uses a period-aware flow matching estimator, multi-period estimator, and discrete wavelet transform to improve performance in Mel-spectrogram reconstruction and text-to-speech tasks.
MohammadReza Ebrahimi, Sunny Panchal, Roland Memisevic
Transformer-based models fail to generalize to longer problem instances due to an inability to randomly access memory within their context window, as demonstrated through analysis of attention maps and examination of effective memory access methodologies.
Carmelo Sferrazza, Dun-Ming Huang, Fangchen Liu +2 authors
The Body Transformer (BoT) architecture, which represents the robot body as a graph and uses masked attention, outperforms vanilla transformers and multilayer perceptrons in task completion, scaling, and computational efficiency for robot learning.
Ruicheng Feng, Chongyi Li, Chen Change Loy
A Kalman-inspired framework named KEEP enhances video face super-resolution by maintaining temporal consistency and capturing facial details across frames.
Kamyar Zeinalipour, Neda Jamshidi, Monica Bianchini +2 authors
Pre-trained LLMs can generate high-quality protein sequences with limited data, achieving performance comparable to specialized protein-focused models.
Rohit Saxena, Frank Keller
A new dataset, MovieSum, is introduced for abstractive summarization of movie screenplays, addressing challenges in processing long contexts and underlining the lack of research in this area compared to television transcripts.
Iretiayo Akinola, Jie Xu, Jan Carius +2 authors
TacSL is a GPU-based library for simulating and learning visuotactile sensor signals, enabling efficient simulation, policy learning, and sim-to-real transfer in contact-rich manipulation tasks.
James Baker
Textual inversion is optimized using vision transformers and bonus tokens to improve adherence to source images and prompts without explicit use of UNet.
Jiawei Huang, Chen Zhang, Yi Ren +6 authors
MulliVC is a novel voice conversion system that separately converts timbre and preserves content and prosody in monolingual and cross-lingual scenarios without requiring multi-lingual paired data.
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