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
Zhuofan Zong, Dongzhi Jiang, Bingqi Ma +5 authors
EasyRef enhances diffusion models for multi-reference image generation by leveraging multimodal large language models and efficient training strategies, outperforming existing methods in aesthetic quality and generalization.
Kaiyi Huang, Yukun Huang, Xuefei Ning +3 authors
GenMAC, a multi-agent framework, enhances text-to-video generation by decomposing complex tasks into simpler stages managed by specialized agents, ensuring high-quality and accurate output through iterative refinement.
Ariel Gera, Odellia Boni, Yotam Perlitz +3 authors
A study evaluates LLM judges for system-level ranking in generative AI, assessing their quality, decisiveness, and bias through comparison with human-based rankings.
Hyeonho Jeong, Chun-Hao Paul Huang, Jong Chul Ye +2 authors
Track4Gen unifies video generation and point tracking to reduce appearance drift, creating temporally stable and visually coherent videos.
Haitam Ben Yahia, Denis Korzhenkov, Ioannis Lelekas +2 authors
A mobile-optimized video diffusion model reduces computational demands and maintains quality through pruning and adversarial finetuning.
Linke Ouyang, Yuan Qu, Hongbin Zhou +17 authors
Document content extraction is crucial in computer vision, especially for meeting the high-quality data needs of large language models (LLMs) and retrieval-augmented generation (RAG) technologies. However, current document parsing methods suffer from significant limitations in terms of diversity and comprehensive evaluation. To address these challenges, we introduce OmniDocBench, a novel multi-source benchmark designed to advance automated document content extraction. OmniDocBench includes a meticulously curated and annotated high-quality evaluation dataset comprising nine diverse document types, such as academic papers, textbooks, slides, among others. Our benchmark provides a flexible and comprehensive evaluation framework with 19 layout category labels and 14 attribute labels, enabling multi-level assessments across entire datasets, individual modules, or specific data types. Using OmniDocBench, we perform an exhaustive comparative analysis of existing modular pipelines and multimodal end-to-end methods, highlighting their limitations in handling document diversity and ensuring fair evaluation. OmniDocBench establishes a robust, diverse, and fair evaluation standard for the document content extraction field, offering crucial insights for future advancements and fostering the development of document parsing technologies. The codes and dataset is available in https://github.com/opendatalab/OmniDocBench.
Tong Wu, Yinghao Xu, Ryan Po +6 authors
A fine-grained visual attribute adaptation framework enables users to selectively apply specific attributes from source images into generated images using a new dataset.
Inkit Padhi, Manish Nagireddy, Giandomenico Cornacchia +19 authors
Granite Guardian models provide comprehensive risk detection for prompts and responses in large language models (LLM), addressing social bias, profanity, violence, sexual content, unethical behavior, jailbreaking, and retrieval-augmented generation (RAG) hallucination risks.
Xiaohui Chen, Satya Narayan Shukla, Mahmoud Azab +8 authors
A new framework called Composite Captions (CompCap) enhances the understanding of composite images by MLLMs through the creation of a large, detailed dataset, leading to significant performance improvements.
Seungtae Nam, Xiangyu Sun, Gyeongjin Kang +3 authors
Generative Densification enhances feed-forward Gaussian models by efficiently upsampling feature representations to improve high-frequency detail representation in 3D reconstruction tasks.
Zexin He, Tengfei Wang, Xin Huang +2 authors
Neural LightRig uses 2D diffusion priors and a U-Net backbone G-buffer model to accurately estimate geometry and materials from a single image under diverse lighting conditions.
Jihao Liu, Zhiding Yu, Shiyi Lan +5 authors
StreamChat enhances Large Multimodal Models' interaction with streaming video by updating visual context at each decoding step and using a crossattention-based architecture with a 3D-RoPE mechanism for temporal information encoding.
Adil Karjauv, Noor Fathima, Ioannis Lelekas +3 authors
Optimizations including a lightweight autoencoder, classifier-free guidance distillation, and an adversarial distillation scheme enable high-speed, high-quality video editing on mobile devices.
Xiao Fu, Xian Liu, Xintao Wang +7 authors
3DTrajMaster enables precise multi-entity 3D motion control in video generation by leveraging 6DoF pose sequences and a 3D-motion grounded object injector with gated self-attention mechanisms.
Namgyu Kang, Jaemin Oh, Youngjoon Hong +1 authors
Physics-Informed Gaussians combine Gaussian feature embeddings with lightweight neural networks to mitigate spectral bias and offer flexible, high-accuracy solutions for Partial Differential Equations.
Pedro H. V. Valois, Lincon S. Souza, Erica K. Shimomoto +1 authors
A framework extending the Linear Representation Hypothesis to multi-token words improves interpretability and control of Large Language Models, demonstrating applications in guided text generation and bias detection.
Mahtab Bigverdi, Zelun Luo, Cheng-Yu Hsieh +4 authors
AURORA, a method that incorporates Perception Tokens into MLMs, improves visual reasoning by transforming intermediate image representations into tokenized formats, enhancing performance on tasks like counting and depth estimation.
Songlin Yang, Jan Kautz, Ali Hatamizadeh
Gated DeltaNet, combining gating and delta update mechanisms, outperforms existing models in various tasks and achieves better performance through hybrid architectures with additional attention layers.
Jixuan Fan, Wanhua Li, Yifei Han +1 authors
A novel momentum-based self-distillation method for 3D Gaussian Splatting enhances reconstruction accuracy and efficiency in large-scale scene reconstruction by addressing memory and storage challenges.
Xiao Xu, Tianhao Niu, Yuxi Xie +3 authors
Integrating fine-grained concept annotations enhances multimodal large language models, improving their performance in vision-language tasks by aligning vision and language at multiple granularities.
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