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。
Zichen Liu, Yue Yu, Hao Ouyang +6 authors
MagicQuill integrates an interface with a multimodal large language model and diffusion prior to enable efficient and precise real-time image editing through minimal user input.
48 篇论文 · 按点赞排序
Zhengyi Wang, Jonathan Lorraine, Yikai Wang +4 authors
The work demonstrates the capability of LLMs to generate 3D meshes from text by introducing a novel approach to tokenize 3D mesh data, allowing the unification of 3D and text modalities without expanding the model's vocabulary.
Yoad Tewel, Rinon Gal, Dvir Samuel Yuval Atzmon +2 authors
Add-it extends diffusion models with a weighted extended-attention mechanism to seamlessly integrate objects into images based on text instructions, achieving state-of-the-art results on real and generated image insertion benchmarks.
Siheng Li, Cheng Yang, Zesen Cheng +4 authors
A novel method enables LLMs to self-improve in long-context reasoning through sampling, scoring with Minimum Bayes Risk, and supervised fine-tuning, outperforming expert-annotated data approaches.
Erik Wijmans, Brody Huval, Alexander Hertzberg +2 authors
Cut Cross-Entropy reduces the memory footprint of large language models during training by efficiently computing cross-entropy loss without materializing logits for all tokens.
Cong Wei, Zheyang Xiong, Weiming Ren +3 authors
An omnipotent image editor trained on multisource supervision, importance sampling, and a versatile architecture excels in multiple tasks with diverse aspect ratios and high data quality.
Yew Ken Chia, Liying Cheng, Hou Pong Chan +5 authors
A benchmark and framework named M-LongDoc evaluates and improves large multimodal models for document reading, focusing on open-ended solutions and retrieval tasks over lengthy documents.
Zhangchen Xu, Fengqing Jiang, Luyao Niu +2 authors
The Larger Models' Paradox reveals that larger models are not always better teachers for fine-tuning smaller models, and a new metric, Compatibility-Adjusted Reward (CAR), is introduced to measure and improve the effectiveness of response generators.
Weiquan Huang, Aoqi Wu, Yifan Yang +8 authors
LLM2CLIP leverages large language models to enhance CLIP's multimodal representation learning, enabling better handling of complex image captions by fine-tuning the LLM and using contrastive learning.
Haolin Chen, Yihao Feng, Zuxin Liu +8 authors
LaTent Reasoning Optimization (LaTRO) enhances LLMs' reasoning capabilities through variational optimization of latent distributions, improving zero-shot accuracy on complex reasoning tasks.
Yancheng He, Shilong Li, Jiaheng Liu +11 authors
Chinese SimpleQA is a comprehensive benchmark focusing on evaluating the factuality of LLMs for Chinese short questions through diverse and high-quality data.
Yiyang Ma, Xingchao Liu, Xiaokang Chen +11 authors
JanusFlow unifies image understanding and generation using rectified flow within a language model framework, demonstrating superior performance across benchmarks.
NVIDIA, Yuval Atzmon, Maciej Bala +28 authors
Edify Image, a family of diffusion models using Laplacian diffusion, generates photorealistic images for various applications with high accuracy.
Yunhan Yang, Yukun Huang, Yuan-Chen Guo +5 authors
SAMPart3D is a scalable and flexible zero-shot 3D part segmentation framework that leverages text-agnostic vision foundation models and multi-view renderings for semantic labeling.
Xiaofeng Wang, Kang Zhao, Feng Liu +6 authors
EgoDreamer generates egocentric videos using EgoVid-5M, a new high-quality dataset with detailed action annotations and kinematic controls, enhancing virtual reality, augmented reality, and gaming applications.
Anas Awadalla, Le Xue, Manli Shu +13 authors
A dataset of 218 million image-text pairs bridges synthetic and factual captions, enhancing vision-language models.
Tom Sander, Pierre Fernandez, Alain Durmus +2 authors
The Watermark Anything Model (WAM) is a deep-learning approach for localized image watermarking, offering high imperceptibility and robustness, including capabilities to locate and extract distinct watermarks from small image regions.
Nizar Islah, Justine Gehring, Diganta Misra +4 authors
A new dataset, GitChameleon, evaluates large language models' ability to generate version-specific and executable code, highlighting their limitations.
Xinghua Zhang, Haiyang Yu, Cheng Fu +2 authors
TRACE is a benchmark for improving complex instruction-following in LLMs, using IOPO to enhance input-output alignment, demonstrating superior performance over SFT and DPO.
Jing Xiong, Gongye Liu, Lun Huang +17 authors
This survey explores the application of autoregressive models in computer vision, categorizing them into pixel-based, token-based, and scale-based models, and discussing their integration with other generative models and applications in emerging domains.
Man Tsung Yeung, Penghui Qi, Min Lin +1 authors
Proposed techniques achieve balanced computation and memory usage in pipeline parallelism for transformer-based models by partitioning vocabulary layers and optimizing communication barriers, resulting in improved throughput and reduced peak memory usage.
Canyu Chen, Jian Yu, Shan Chen +5 authors
Despite their capabilities in text processing, existing large language models do not currently outperform traditional machine learning models in clinical prediction tasks.
Qingyu Yin, Chak Tou Leong, Hongbo Zhang +8 authors
A novel method, Feature-level constrained Preference Optimization (FPO), improves the alignment of large language models with human preferences by using sparse features from Sparse Autoencoders and feature-level offline reference, achieving better efficiency and win rate.
Yuze He, Yanning Zhou, Wang Zhao +5 authors
StdGEN generates high-quality, semantically decomposed 3D characters from single images using a transformer-based model and efficient diffusion techniques, outperforming existing methods in speed and quality.
Wissam Antoun, Francis Kulumba, Rian Touchent +3 authors
New versions of CamemBERT, CamemBERTav2 and CamemBERTv2, address temporal concept drift using DeBERTaV3 and RoBERTa architectures, respectively, outperforming their predecessors across various NLP tasks.
Rahul Ravishankar, Zeeshan Patel, Jathushan Rajasegaran +1 authors
Diffusion models enhance visual perception tasks like depth estimation, optical flow, and segmentation through efficient training and reduced data and compute requirements.
Noël Vouitsis, Rasa Hosseinzadeh, Brendan Leigh Ross +4 authors
Direct Consistency Models aim to reduce error in solving the probability flow ODE of diffusion models but surprisingly degrade sample quality compared to traditional Consistency Models.
Aditya Sanghi, Aliasghar Khani, Pradyumna Reddy +5 authors
Wavelet Latent Diffusion efficiently generates high-quality 3D shapes with high compression and minimal detail loss, improving performance and computational efficiency.
Jingwei Xu, Chenyu Wang, Zibo Zhao +3 authors
A unified CAD generation system, CAD-MLLM, uses advanced large language models to align multimodal data and generates high-quality parametric CAD models conditioned on textual, image, and point cloud inputs.
Ishika Agarwal, Krishna Killamsetty, Lucian Popa +1 authors
DELIFT optimizes data selection during LLM fine-tuning across multiple stages using a pairwise utility metric, reducing data size and computational costs while maintaining high performance.
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