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
Shijie Zhou, Zhiwen Fan, Dejia Xu +6 authors
A text-to-3D 360° scene generation pipeline creates immersive environments using a 2D diffusion model and lifts flat panoramas into 3D Gaussians with semantic and geometric constraints.
Robert Vacareanu, Vlad-Andrei Negru, Vasile Suciu +1 authors
Large language models can perform regression tasks competitively with traditional supervised methods and achieve sub-linear regret with increasing in-context examples.
Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar +3 authors
BRAVE consolidates features from multiple vision encoders to improve VLM performance on captioning and VQA tasks, addressing visual encoding limitations with reduced parameters and compressed representation.
Jiahao Wang, Wenqi Shao, Mengzhao Chen +6 authors
A decoder-only Transformer model, LLaMA, is adapted for computer vision tasks through architectural modifications, achieving performance comparable to encoder-only models with fewer parameters and improved attention mechanisms.
Zifeng Wang, Chun-Liang Li, Vincent Perot +5 authors
CodecLM is a framework that utilizes LLMs to generate high-quality, aligned synthetic data for instruction-following tasks, improving adaptability and performance across various LLMs and instruction distributions.
Yang Zheng, Qingqing Zhao, Guandao Yang +8 authors
PhysAvatar integrates inverse rendering and physics simulation to reconstruct and render realistic avatars in various dynamic and lighting conditions using multi-view video data.
Sergey Kastryulin, Artem Konev, Alexander Shishenya +20 authors
YaART, a text-to-image cascaded diffusion model enhanced with RLHF, achieves high-quality image generation efficiently and outperforms existing state-of-the-art models in user evaluations.
Xingwei Qu, Yuelin Bai, Yinghao Ma +26 authors
Large Language Models are adapted for music pre-training using ABC Notation instead of MIDI, leading to enhanced musical composition through synchronized multi-track handling and exploration of the Symbolic Music Scaling Law.
Arushi Goel, Zhifeng Kong, Rafael Valle +1 authors
A new multi-turn dialogue dataset for audio, Audio Dialogues, is introduced to enhance understanding and dialogue capabilities in audio processing using a Large Language Model.
Diyi Yang, Caleb Ziems, William Held +3 authors
This perspective paper proposes using large language models to create an AI Partner and AI Mentor framework for social skill training, integrating experiential learning with personalized feedback to address barriers in workforce development and social equality.
Owen Oertell, Jonathan D. Chang, Yiyi Zhang +2 authors
A reinforcement learning approach for fine-tuning consistency models enhances text-to-image generation by improving quality and reducing inference time compared to diffusion models.
Tuomas Kynkäänniemi, Miika Aittala, Tero Karras +3 authors
Applying guidance only within a specific noise level interval in image-generating diffusion models improves inference speed and result quality across various settings.
Kunpeng Song, Yizhe Zhu, Bingchen Liu +3 authors
MoMA, an open-vocabulary and training-free model, uses a Multimodal Large Language Model to generate high-detail, identity-preserved images through an image diffusion model enhanced by a self-attention shortcut.
Shufan Li, Konstantinos Kallidromitis, Akash Gokul +2 authors
Diffusion-KTO fine-tunes text-to-image diffusion models using per-image binary feedback, enhancing performance in both human and automatic evaluations.
Ji Liu, Zifeng Zhang, Mingjie Lu +7 authors
A transformer-based lane detection framework using sparse anchors outperforms state-of-the-art methods with fewer computations.
Zoltan Csaki, Bo Li, Jonathan Li +7 authors
This investigation explores comprehensive techniques for adapting large language models to new languages, addressing vocabulary extension, preference optimization, and data scarcity issues across multiple languages and model sizes.
Xinrun Du, Zhouliang Yu, Songyang Gao +13 authors
CT-LLM, a large language model emphasizing Chinese language training, demonstrates exceptional proficiency through extensive Chinese textual data and outperforms in both Chinese and English tasks.
Jingxuan Xu, Wuyang Chen, Yao Zhao +1 authors
The approach leverages smaller vision-language models to achieve efficient Open-Vocabulary Segmentation with comparable or better performance and lower computational costs compared to larger models.
Xingyi Yang, Xinchao Wang
Hash3D accelerates 3D generation by hashing and reusing feature maps, enhancing smoothness and reducing processing time for text-to-3D and image-to-3D models.
Zhengcong Fei, Mingyuan Fan, Changqian Yu +2 authors
Diffusion-RWKV, an architecture adapted from RWKV for image generation, efficiently handles high-resolution images with reduced computational complexity compared to CNN or Transformer-based models.
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