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
Jan-Niklas Dihlmann, Andreas Engelhardt, Hendrik Lensch
SIGNeRF, a novel method for NeRF scene editing and object generation, utilizes depth-conditioned diffusion models for 3D consistent image updates without iterative optimization.
Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios +5 authors
ODIN is a transformer-based model that segments both 2D and 3D data by alternating between 2D and 3D information fusion, achieving state-of-the-art performance on multiple benchmarks.
Tao He, Xue Li, Zhibin Wang +4 authors
Unicron, a workload manager, enhances the self-healing and efficiency of large-scale language model training by optimizing for cost and minimizing downtime across multiple tasks.
Uri Shaham, Jonathan Herzig, Roee Aharoni +3 authors
Multilingual tuning of large language models with a small number of multilingual examples improves instruction-following capabilities across languages compared to monolingual tuning.
Yifang Men, Biwen Lei, Yuan Yao +3 authors
En3D generates high-quality 3D human avatars using synthetic 2D data without relying on pre-existing 3D models, employing a combination of a 3D generator, geometry sculptor, and texturing module.
Hongkun Hao, Long Zhou, Shujie Liu +4 authors
Coupling large language models with text-to-speech synthesis by using the LLM as a text encoder significantly improves speech quality and reduces word error rate.
Zizhang Li, Dor Litvak, Ruining Li +6 authors
3D-Fauna learns a deformable 3D animal model using 2D images and the Semantic Bank of Skinned Models to handle diverse and rare animal species efficiently.
Chen Zheng, Ke Sun, Da Tang +4 authors
ICE-GRT, a Reinforcement Learning from Human Feedback approach using Proximal Policy Optimization, enhances domain-specific performance while maintaining general task capabilities in Large Language Models.
Yiwen Lu, Zhen Ye, Wei Xue +3 authors
CoMoSVC is a consistency model-based singing voice conversion method that achieves fast inference speed and high-quality audio generation.
Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad +5 authors
LLMs can learn aspects of the visual world through the precise modeling of string representations of images, which can also aid in training vision models for semantic assessments of natural images.
Muyang Du, Chuan Liu, Junjie Lai
Incremental FastPitch improves real-time speech synthesis by enabling incremental Mel chunk generation with reduced latency.
Peihao Wang, Dejia Xu, Zhiwen Fan +8 authors
Maximizing entropy in score distillation improves text-to-3D generation by addressing view inconsistency issues and mode collapse via a method called Entropic Score Distillation.
Xiaotong Wu, Wei-Sheng Lai, YiChang Shih +4 authors
An efficient hybrid zoom super-resolution system for mobile devices uses machine learning to align and transfer details from telephoto to wide shots, addressing depth-of-field mismatches and other issues.
Chunyi Li, Haoning Wu, Zicheng Zhang +7 authors
Q-Refine, a quality-award refiner guided by IQA metrics, effectively optimizes AI-generated images across different qualities and models, enhancing both fidelity and aesthetic quality.
Zhouhan Lin, Cheng Deng, Le Zhou +20 authors
GeoGalactica, a specialized LLM for geoscience with 30 billion parameters, was created through further pre-training and supervised fine-tuning with a large geoscience-related text corpus and professional instruction-tuning data.
Ling Yang, Jingwei Liu, Shenda Hong +5 authors
ConPreDiff enhances diffusion models by incorporating context prediction to improve image synthesis quality across various tasks.
Xingxing Zuo, Pouya Samangouei, Yunwen Zhou +2 authors
The work introduces a method to integrate vision-language embeddings into 3D Gaussian Splatting for efficient 3D reconstruction, improving multi-view semantic consistency and inference speed.
Jun-Yan He, Zhi-Qi Cheng, Chenyang Li +10 authors
The WordArt Designer API uses Large Language Models to enable intuitive, user-driven creation of unique typographic designs, enhancing user satisfaction and flexibility over traditional methods.
Aleksandar Stanić, Sergi Caelles, Michael Tschannen
A framework leveraging abstract routines and automated in-context example generation enhances visual reasoning tasks, improving performance and reducing reliance on human engineering for large language models.
Peihao Wang, Zhiwen Fan, Dejia Xu +8 authors
Stein Score Distillation (SSD) reduces variance in score distillation for text-to-3D asset synthesis, improving visual quality and convergence speed.
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