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
Takahide Yoshida, Atsushi Masumori, Takashi Ikegami
A humanoid robot, Alter3, uses a Large Language Model to generate spontaneous and varied body movements, demonstrating zero-shot learning and adjustability through verbal feedback.
Yuanyou Xu, Zongxin Yang, Yi Yang
SEEAvatar generates photorealistic 3D avatars from text using a self-evolving template for geometry and a diffusion model-enhanced rendering pipeline for appearance.
Kate Baumli, Satinder Baveja, Feryal Behbahani +23 authors
Researchers explore using off-the-shelf vision-language models to derive rewards for reinforcement learning agents, demonstrating improved performance across various language-based visual goals.
Kaiqiang Song, Xiaoyang Wang, Sangwoo Cho +2 authors
A new model architecture called Zebra addresses the challenges of handling large text sequences in LLMs by using grouped local-global attention layers, improving efficiency and performance.
Amirhossein Habibian, Amir Ghodrati, Noor Fathima +4 authors
Clockwork Diffusion reduces computational complexity in text-to-image diffusion models by reusing low-resolution feature maps, maintaining perceptual quality.
Bowen Wen, Wei Yang, Jan Kautz +1 authors
FoundationPose, a unified foundation model using neural implicit representation and transformer architecture, outperforms specialized models in 6D object pose estimation and tracking.
Nina Rimsky, Nick Gabrieli, Julian Schulz +3 authors
Contrastive Activation Addition (CAA) modifies model activations to steer language model behavior with high precision and insight into high-level concept representation.
Jie Xiao, Kai Zhu, Han Zhang +5 authors
Research explores methods to integrate ControlNet-like conditional control into Consistency Models, finding approaches for high-level semantic control, alternative training methods, and lightweight adapters.
Yang Fu, Sifei Liu, Amey Kulkarni +3 authors
The paper presents a method for novel view synthesis and camera pose estimation using 3D Gaussian Splatting, eliminating the need for pre-processed camera poses by processing input frames sequentially.
Junbum Cha, Wooyoung Kang, Jonghwan Mun +1 authors
A novel visual projector for Multimodal Large Language Models enhances flexibility and locality, leading to improved performance across various benchmarks.
Yuewei Yang, Xiaoliang Dai, Jialiang Wang +2 authors
A quantization strategy leveraging global and local treatments efficiently compresses Latent Diffusion Models (LDMs) for edge deployment, maintaining model performance.
Joanna Materzynska, Josef Sivic, Eli Shechtman +3 authors
The method fine-tunes text-to-video models to incorporate customized motions from a few video samples, supports multi-person and multimodal customization, and includes a quantitative evaluation approach.
Hongsuk Choi, Isaac Kasahara, Selim Engin +3 authors
FineControlNet enhances text-driven image generation by allowing fine control over instance appearance and maintaining pose control, outperforming existing methods in generating images that align with user-provided prompts and poses.
Xueyan Zou, Linjie Li, Jianfeng Wang +9 authors
FIND is a transformer interface that aligns foundation model embeddings for unified image segmentation and dataset-level retrieval without tuning the foundation models.
Zhen Wang, Qiangeng Xu, Feitong Tan +6 authors
MVDD is a denoising diffusion model using multi-view depth representations to generate high-quality 3D shape point clouds and meshes, achieving state-of-the-art results in 3D shape generation and depth completion.
Olga Golovneva, Sean O'Brien, Ramakanth Pasunuru +4 authors
PathFinder, a tree-search-based reasoning model, improves performance on multi-step reasoning tasks by integrating dynamic decoding, constraints, pruning, and ranking.
Enis Simsar, Alessio Tonioni, Yongqin Xian +2 authors
LIME enhances localized image editing in diffusion models using pre-trained features, clustering, refined cross-attention, and regularization without re-training or fine-tuning.
Junfeng Wu, Yi Jiang, Qihao Liu +3 authors
GLEE is an object-level foundation model for tasks like detection, segmentation, tracking, and identification, demonstrating zero-shot transfer and state-of-the-art performance.
Xiaoxia Wu, Haojun Xia, Stephen Youn +8 authors
FP6 quantization offers robust performance across various tasks in large language models, surpassing INT4 quantization through a novel 4+2 design that maintains latency.
Hakan Inan, Kartikeya Upasani, Jianfeng Chi +8 authors
Llama Guard, an LLM-based safeguard model, uses a safety risk taxonomy for prompt and response classification, demonstrating strong performance on content moderation tasks with instruction fine-tuning.
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