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。
43 篇论文 · 按点赞排序
Michael Luo, Justin Wong, Brandon Trabucco +5 authors
Stylus automatically selects and composes task-specific adapters based on prompts, improving the efficiency and quality of custom image generation.
Evangelos Ververas, Rolandos Alexandros Potamias, Jifei Song +2 authors
A structure-aware Gaussian Splatting method improves rendering performance and reduces storage requirements by encoding scene geometry without compression.
Yasumasa Onoe, Sunayana Rane, Zachary Berger +9 authors
The Descriptions of Connected and Contrasting Images (DOCCI) dataset provides detailed human-annotated descriptions for image-to-text and text-to-image generation, demonstrating effectiveness in training and highlighting limitations in current models.
Puhao Li, Tengyu Liu, Yuyang Li +6 authors
Ag2Manip addresses domain gaps and sparsity in robotic manipulation tasks by introducing agent-agnostic visual and action representations, achieving significant performance improvements in both simulation and real-world settings.
Shangzhan Zhang, Sida Peng, Tao Xu +7 authors
A method generates materials for 3D meshes from text by using segment-wise procedural material graphs and a diffusion model bridge.
Paul Engstler, Andrea Vedaldi, Iro Laina +1 authors
Proposes a new depth completion model and benchmarking scheme for 3D scene generation, improving geometric coherence and structural quality.
Kai He, Kaixin Yao, Qixuan Zhang +3 authors
A text-driven 3D garment generation framework, DressCode, uses SewingGPT and Stable Diffusion to create high-quality, CG-friendly garments through natural language interaction.
Runyi Yang, Zhenxin Zhu, Zhou Jiang +6 authors
SUNDAE, a memory-efficient 3D Gaussian field, reduces memory usage by pruning primitives using spectral down-sampling and compensates for quality loss with a neural network head, achieving comparable or better performance than standard Gaussian splatting.
Luxi Chen, Zhengyi Wang, Chongxuan Li +3 authors
Score-based iterative reconstruction (SIR) enhances zero-shot 3D generation efficiency compared to score distillation sampling by optimizing 3D parameters iteratively, reducing the number of function evaluations and achieving faster generation times.
Longlong Jing, Ruichi Yu, Xu Chen +19 authors
STT, a Transformer-based model, effectively tracks objects and estimates their states by optimizing both association and state estimation tasks with enhanced metrics.
Alex Kim, Jia Huang, Rob Monarch +5 authors
A novel image-text matching model fine-tuned on LXMERT improves the accuracy of matching app images to search terms compared to CLIP and a transformer-ResNet baseline.
Chen Dudai, Morris Alper, Hana Bezalel +3 authors
A localization system using vision-and-language models enhances understanding of architectural landmarks by integrating textual and image data, improving localization accuracy over other 3D and 2D models.
Ang Cao, Justin Johnson, Andrea Vedaldi +1 authors
Lightplane Render and Splatter components reduce memory usage in 2D-3D mapping for 3D neural fields, enabling efficient processing and scalable reconstruction and generation.
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