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。
47 篇论文 · 按点赞排序
Tzofi Klinghoffer, Xiaoyu Xiang, Siddharth Somasundaram +4 authors
A single-photon lidar-based approach combines NeRF and two-bounce light measurements to achieve accurate single-view 3D reconstruction without data priors or controlled lighting.
Tycho F. A. van der Ouderaa, Markus Nagel, Mart van Baalen +2 authors
Data-driven compression using Kronecker-factored curvature approximations efficiently prunes large language models with minimal performance impact.
Mengyao Lyu, Yuhong Yang, Haiwen Hong +6 authors
A novel erasing framework using one-dimensional adapters and Latent Anchoring fine-tuning strategy mitigates undesired concept alternations and enhances transferability in diffusion models for text-to-image generation.
Linyi Yang, Shuibai Zhang, Zhuohao Yu +8 authors
A framework using task-specific fine-tuning improves large language models' generalizability and factuality during inference by leveraging discriminative models.
Towaki Takikawa, Thomas Müller, Merlin Nimier-David +4 authors
Hash tables with learned probes achieve high quality and fast inference for neural graphics primitives without the drawbacks of existing feature grids.
Christian Simon, Sen He, Juan-Manuel Perez-Rua +3 authors
A neural rendering technique using signed distance functions, geometry-encoding volumes, and HyperNetworks achieves scene generalization and consistency without scene-specific optimization.
Zixuan Huang, Stefan Stojanov, Anh Thai +2 authors
A regression-based model named ZeroShape achieves superior performance and higher computational efficiency in single-image zero-shot 3D shape reconstruction compared to generative models.
Yuxuan Zhang, Jiaming Liu, Yiren Song +8 authors
The SSR-Encoder is a novel architecture that selectively captures and focuses on crucial subject representations from reference images, enhancing subject-driven image generation without test-time fine-tuning.
Guansong Lu, Yuanfan Guo, Jianhua Han +7 authors
PanGu-Draw, a latent diffusion model, enhances resource efficiency in text-to-image synthesis by integrating variable control signals through time-decoupling and cooperative diffusion techniques.
Katrin Tomanek, Shanqing Cai, Subhashini Venugopalan
Personalizing a large language model's suggestions through fine-tuning, prompt-tuning, and retrieval augmentation enhances relevance, with prompt-tuning and retrieval augmentation outperforming fine-tuning in few-shot scenarios.
Biwen Lei, Kai Yu, Mengyang Feng +2 authors
A new framework, DiffusionGAN3D, enhances 3D domain adaptation and avatar generation by combining 3D GANs with diffusion priors, improving quality and efficiency through relative distance loss and a progressive texture refinement module.
Alicia Golden, Samuel Hsia, Fei Sun +8 authors
The study characterizes performance issues in Diffusion- and Transformer-based text-to-image and text-to-video models, revealing specific breakdowns and optimization opportunities for these multi-modal workloads.
Siddhartha Datta, Alexander Ku, Deepak Ramachandran +1 authors
A Prompt Expansion framework optimizes text queries to generate diverse and high-quality images through text-to-image models.
Soshi Shimada, Franziska Mueller, Jan Bednarik +7 authors
MACS, using cascaded diffusion models, synthesizes realistic 3D hand-object interactions conditioned on mass, supporting various applications like training data generation and computer games.
Zhongyang Guo, Guanran Jiang, Zhongdan Zhang +3 authors
Shai, a 10B-level large language model for asset management, demonstrates enhanced performance through continuous pre-training and fine-tuning, using a comprehensive evaluation framework that includes professional qualification exams and safety assessments.
Zheng Ding, Xuaner Zhang, Zhuowen Tu +1 authors
A method for image restoration using denoising diffusion models with constraints from anchor images to preserve identity and high-frequency details is proposed, showing superior performance compared to existing methods.
Wei Feng, Xin Wang, Hong Chen +4 authors
The LLM4VG benchmark evaluates video grounding performance of LLMs, finding VidLLMs insufficient and combinations of LLMs with visual models showing potential improvement.
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