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
Kaixiong Gong, Kaituo Feng, Bohao Li +8 authors
AV-Odyssey Bench evaluates the audio-visual understanding of large language models through a comprehensive benchmark of 4,555 multiple-choice questions.
Jun Xiang, Yudong Guo, Leipeng Hu +3 authors
A new method creates photorealistic, expressive whole-body talking avatars from a single image using pose-guided image-to-video diffusion models and a hybrid 3DGS-mesh representation with regularizations.
Junyuan Zhang, Qintong Zhang, Bin Wang +6 authors
Introducing OHRBench to evaluate OCR's impact on Retrieval-augmented Generation (RAG) systems and explore the use of Vision-Language Models (VLMs) as an alternative to OCR.
Junha Hyung, Kinam Kim, Susung Hong +2 authors
Spatiotemporal Skip Guidance enhances video diffusion models by using layer skipping to achieve high sample quality without compromising diversity or dynamics.
Zehuan Huang, Yuan-Chen Guo, Haoran Wang +4 authors
MV-Adapter, an adapter-based solution, enhances text-to-image models for multi-view image generation with efficient training, preserving pre-trained knowledge and facilitating the integration of 3D data.
Jianping Jiang, Weiye Xiao, Zhengyu Lin +7 authors
SOLAMI is an end-to-end VLA framework for 3D autonomous characters that addresses social interaction by generating multimodal responses and enabling immersive VR interaction.
Justin Chih-Yao Chen, Zifeng Wang, Hamid Palangi +8 authors
RevThink, a framework combining data augmentation and multi-task learning, enhances reverse thinking in Large Language Models, leading to improved zero-shot performance and generalization.
Xinghui Li, Qichao Sun, Pengze Zhang +5 authors
AnyDressing enhances garment-centric image generation by integrating garment-specific features and personalized text prompts using GarmentsNet and DressingNet, achieving state-of-the-art results in multi-garment virtual dressing.
Jaskirat Singh, Lindsey Li, Weijia Shi +7 authors
Adversarial guidance using visual features from reference images enhances diversity and reduces visual similarity to copyrighted assets during reverse diffusion processes.
Jeongho Ju, Daeyoung Kim, SunYoung Park +1 authors
An open-source Korean-English vision-language model, VARCO-VISION, is introduced and demonstrates superior performance in bilingual image-text tasks compared to similar models, incorporating a step-by-step training strategy and supporting grounding, referring, and OCR.
Akhiad Bercovich, Tomer Ronen, Talor Abramovich +21 authors
Puzzle is a framework that optimizes large language model inference using neural architecture search, blockwise local knowledge distillation, and mixed-integer programming to achieve faster throughput and reduced computational costs while preserving model accuracy.
Kangsan Kim, Geon Park, Youngwan Lee +2 authors
VideoICL enhances out-of-distribution video understanding with similarity-based example selection and iterative inference, extending context length without high computational costs.
Zehuan Huang, Yuan-Chen Guo, Xingqiao An +7 authors
MIDI is a novel method for 3D scene generation from a single image using multi-instance diffusion models with attention mechanisms, achieving state-of-the-art results across different datasets.
Shivalika Singh, Angelika Romanou, Clémentine Fourrier +20 authors
Global-MMLU addresses cultural biases and translation artifacts in multilingual datasets by involving professional annotators and offering culturally distinct subsets for more accurate model evaluation.
Lingen Li, Zhaoyang Zhang, Yaowei Li +7 authors
NVComposer, a novel generative approach, achieves state-of-the-art performance in novel view synthesis by implicitly inferring spatial and geometric relationships between multiple views without requiring explicit external alignment.
Alessandro Scirè, Andrei Stefan Bejgu, Simone Tedeschi +3 authors
LLM-Oasis, a comprehensive dataset for factuality evaluation, challenges state-of-the-art LLMs by requiring high accuracy in distinguishing factual from unfactual texts.
Jinyuan Qu, Hongyang Li, Shilong Liu +3 authors
TAPTRv3 enhances point tracking robustness in long videos by integrating spatial and temporal context through Context-aware Cross-Attention and Visibility-aware Long-Temporal Attention, achieving superior performance compared to existing methods.
Feng Liu, Shiwei Zhang, Xiaofeng Wang +6 authors
A new training-free caching method called TeaCache improves inference speed in diffusion models for video generation without significantly compromising visual quality.
Chaojun Xiao, Jie Cai, Weilin Zhao +4 authors
The paper introduces 'capacity density' as a metric to evaluate the efficiency and effectiveness of Large Language Models and reveals an empirical law showing their capacity density doubles every three months.
Dar-Yen Chen, Hmrishav Bandyopadhyay, Kai Zou +1 authors
NitroFusion uses a dynamic adversarial framework with specialized discriminator groups for high-quality, high-fidelity single-step image generation with flexible quality-speed trade-offs.
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