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
Stanislaw Szymanowicz, Christian Rupprecht, Andrea Vedaldi
Splatter Image, a learning-based method using Gaussian Splatting and cross-view attention, achieves ultra-fast monocular 3D object reconstruction with high-quality results.
Mosam Dabhi, Laszlo A. Jeni, Simon Lucey
A 3D Lifting Foundation Model (3D-LFM) uses permutation-equivariant transformers to perform 2D-3D lifting for various object categories, achieving state-of-the-art performance.
Jie Ren, Yao Zhao, Tu Vu +2 authors
Self-evaluation at the token level improves accuracy and quality in selective generation tasks for large language models.
Desai Xie, Jiahao Li, Hao Tan +6 authors
Carve3D, integrating RLFT with a Multi-view Reconstruction Consistency (MRC) metric, enhances multi-view consistency and NeRF reconstruction quality while mitigating distribution shift in text-to-3D models.
Han Shu, Wenshuo Li, Yehui Tang +5 authors
TinySAM reduces computational cost while maintaining strong zero-shot segmentation performance through knowledge distillation, quantization, and hierarchical segmentation strategies.
Chaoyou Fu, Renrui Zhang, Haojia Lin +19 authors
Preliminary evaluation of Gemini Pro's visual understanding shows it to be a strong contender to GPT-4V in multi-modal learning, though both models face challenges in achieving domain generalization compared to open-sourced systems.
Bingyuan Wang, Hengyu Meng, Zeyu Cai +4 authors
MagicScroll is a diffusion-based image generation framework that ensures coherence and style consistency across diverse aspect ratios for visual storytelling, leveraging text, image, and layout conditions.
Weizhi Fei, Xueyan Niu, Pingyi Zhou +4 authors
A semantic compression method extends the context window of Transformer-based LLMs for longer texts without fine-tuning, maintaining fluency and reducing computational costs.
Senmao Li, Taihang Hu, Fahad Shahbaz Khan +5 authors
Introducing an encoder propagation scheme accelerates diffusion sampling without distilling knowledge, yielding faster and high-quality image and video generation.
Zhaoyang Zhang, Wenqi Shao, Yixiao Ge +3 authors
A Cached Transformer model with Gated Recurrent Cached attention achieves significant advancements in language and vision tasks by extending self-attention with a differentiable memory cache.
Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges +2 authors
HAAR is a text-guided generative model using 3D hair strands and latent diffusion trained in UV space, addressing limitations of 2D-based 3D hairstyle generation.
Shraman Pramanick, Guangxing Han, Rui Hou +6 authors
VistaLLM is a unified framework for vision-language tasks that uses instruction-guided tokenization and adaptive sampling to process coarse- and fine-grained visual inputs with state-of-the-art results.
Cheng-Yen Hsieh, Tarasha Khurana, Achal Dave +1 authors
A new TAO-Amodal benchmark and an amodal expander module improve amodal detection and tracking of occluded objects in video sequences.
Ian Magnusson, Akshita Bhagia, Valentin Hofmann +13 authors
Paloma analyzes the performance of language models across 585 diverse text domains to provide a comprehensive and comparable assessment of their domain fit and cost-effectiveness.
Sudharshan Suresh, Haozhi Qi, Tingfan Wu +9 authors
NeuralFeels, a multimodal perception method combining vision and touch, enhances spatial awareness and object tracking during in-hand manipulation, especially under visual occlusion.
Chaojian Li, Bichen Wu, Peter Vajda +2 authors
MixRT, a novel NeRF representation combining a low-quality mesh, view-dependent displacement map, and compressed NeRF model, achieves real-time rendering on edge devices with improved quality and reduced storage.
Weijia Mao, Yan-Pei Cao, Jia-Wei Liu +2 authors
ShowRoom3D generates high-quality 3D room-scale scenes from texts using a 3D diffusion prior and a progressive view selection process, improving structural integrity and clarity across different perspectives.
Huan Ling, Seung Wook Kim, Antonio Torralba +2 authors
A text-guided compositional generation method uses score distillation and dynamic 3D Gaussian Splatting to synthesize high-quality, animated 4D objects with improved temporal consistency and realism.
Haoyu Ma, Shahin Mahdizadehaghdam, Bichen Wu +5 authors
The study introduces MaskINT, a non-autoregressive masked generative transformer, to efficiently edit videos based on text prompts by combining it with an existing text-to-image diffusion model for keyframe editing.
Qi Mao, Lan Chen, Yuchao Gu +2 authors
MAG-Edit enhances localized image editing in complex scenarios by optimizing noise latent features in diffusion models through mask-based cross-attention constraints.
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