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。
Andreas Steiner, André Susano Pinto, Michael Tschannen +15 authors
PaliGemma 2 integrates a SigLIP-So400m vision encoder with Gemma 2 models of varying sizes and resolutions, advancing transfer performance across diverse vision-language tasks, including OCR and captioning.
50 篇论文 · 按点赞排序
Senqiao Yang, Yukang Chen, Zhuotao Tian +4 authors
VisionZip reduces visual token redundancy and enhances efficiency in vision-language models, improving performance and inference speed across various tasks.
Viet Nguyen, Anh Aengus Nguyen, Trung Dao +4 authors
SNOOPI framework improves one-step text-to-image diffusion models with Proper Guidance-SwiftBrush and Negative-Away Steer Attention, enhancing both training stability and negative prompt support.
Jianfeng Xiang, Zelong Lv, Sicheng Xu +6 authors
A 3D generation method using a unified SLAT representation and rectified flow transformers achieves high-quality results across different formats and conditions.
Yiheng Xu, Zekun Wang, Junli Wang +6 authors
Aguvis is a unified, vision-based framework for autonomous GUI agents that integrates language grounding and explicit planning, enabling cross-platform generalization and outperforming existing methods.
Jiuhai Chen, Jianwei Yang, Haiping Wu +4 authors
Florence-VL enhances multimodal language models with improved visual features from Florence-2, achieving state-of-the-art results across various benchmarks through depth-breath fusion and a novel training recipe.
Zicheng Lin, Tian Liang, Jiahao Xu +6 authors
The proposed cDPO approach identifies and rewards critical tokens to improve the reasoning accuracy of LLMs by utilizing contrastive estimation between positive and negative models.
Zeyi Sun, Ziyang Chu, Pan Zhang +6 authors
X-Prompt, an auto-regressive large-vision language model, uses in-context learning to achieve competitive performance in seen and unseen image generation tasks.
Zhijian Liu, Ligeng Zhu, Baifeng Shi +24 authors
NVILA, a family of VLMs, optimizes efficiency and accuracy through a scale-then-compress approach, enhancing performance across various benchmarks while reducing computational costs.
Mingzhe Zheng, Yongqi Xu, Haojian Huang +8 authors
VGoT generates high-quality, coherent multi-shot videos by structuring the process into script, keyframe, and shot-level generation, using IP embeddings and smoothing mechanisms for consistency and narrative flow.
Seungone Kim, Juyoung Suk, Xiang Yue +7 authors
AgoraBench evaluates language models' data generation abilities, revealing their distinct strengths and demonstrating various factors that impact effectiveness.
Taekyung Ki, Dongchan Min, Gyoungsu Chae
FLOAT method generates high-quality, temporally consistent, and emotion-enhanced talking portraits using flow matching in a learned motion latent space with a transformer-based vector field predictor.
Sumeet Ramesh Motwani, Chandler Smith, Rocktim Jyoti Das +6 authors
Multi-agent LLM training improves performance on reasoning tasks by assigning specialized roles and utilizing joint outcome-based rewards to enhance collaboration among models.
Zijian Zhang, Kaiyuan Zheng, Zhaorun Chen +6 authors
GRAPE improves vision-language-action models' performance by aligning policies via preference modeling, enhancing generalizability and allowing customization of objectives like safety and efficiency.
Yuxiang Zhang, Shangxi Wu, Yuqi Yang +4 authors
O1-CODER integrates reinforcement learning and Monte Carlo Tree Search to enhance coding capabilities, focusing on pseudocode and full code generation through iterative fine-tuning and standardized code testing.
Bingxin Ke, Dominik Narnhofer, Shengyu Huang +5 authors
RollingDepth transforms a single-image latent diffusion model into a video depth estimator by combining multi-frame depth estimation with optimization-based registration for consistent long-video depth prediction.
Enshen Zhou, Qi Su, Cheng Chi +5 authors
Code-as-Monitor (CaM) uses vision-language models to address both reactive and proactive open-set failure detection in closed-loop robotic systems by formulating them as spatio-temporal constraint satisfaction problems.
Jinyang Wu, Mingkuan Feng, Shuai Zhang +3 authors
HiAR-ICL, a high-level automated reasoning paradigm in In-context Learning (ICL), uses atomic reasoning actions, Monte Carlo Tree Search, and thought cards to improve mathematical reasoning accuracy compared to existing models.
Liao Qu, Huichao Zhang, Yiheng Liu +7 authors
TokenFlow, a dual-codebook image tokenizer, enhances multimodal understanding and generation by decoupling semantic and pixel-level features with shared indices.
Lifan Yuan, Wendi Li, Huayu Chen +6 authors
An implicit process reward model can be trained without additional step labels by parameterizing outcome rewards as log-likelihood ratios, outperforming traditional methods with less data and better generalization.
Anton Voronov, Denis Kuznedelev, Mikhail Khoroshikh +2 authors
Switti, a scale-wise transformer for text-to-image generation, improves convergence and performance through architectural modifications, reduces memory usage, and achieves competitive results compared to diffusion models with significant speed advantages.
Bin Lin, Yunyang Ge, Xinhua Cheng +21 authors
Open-Sora Plan is an open-source project that generates high-resolution, long-duration videos from various inputs using a Wavelet-Flow Variational Autoencoder, Joint Image-Video Skiparse Denoiser, and condition controllers.
Jing Tan, Shuai Yang, Tong Wu +4 authors
Imagine360 generates high-quality, personalized 360° videos with rich motion patterns using dual-branch denoising, antipodal masks, and elevation-aware designs.
Daixuan Cheng, Shaohan Huang, Ziyu Zhu +5 authors
The paper explores the domain adaptation of multimodal large language models through post-training, utilizing a visual instruction synthesizer and single-stage training pipeline to improve performance on specific tasks in domains like biomedicine and food.
Donghoon Ahn, Jiwon Kang, Sanghyun Lee +9 authors
A noise-refining method is introduced to eliminate the need for guidance in diffusion models, enhancing image generation speed and memory efficiency.
Weiming Ren, Huan Yang, Jie Min +2 authors
VISTA, a Video Spatiotemporal Augmentation framework, enhances long-duration and high-resolution video understanding through the creation of synthetic video datasets and finetuning of video LMMs.
01. AI, Alan Wake, Albert Wang +39 authors
Yi-Lightning, a large language model with enhanced Mixture-of-Experts architecture and RAISE safety framework, achieves top-tier performance in specialized categories while being cost-effective and addressing safety issues.
Shengyuan Zhang, An Zhao, Ling Yang +7 authors
A novel distillation method, ScoreLiDAR, accelerates 3D LiDAR scene completion with high-quality results by reducing sampling steps and incorporating a Structural Loss to maintain geometric structure.
Yiwu Zhong, Zhuoming Liu, Yin Li +1 authors
An adaptive inference method reduces computational demands in multi-modal LLMs for video and image understanding without significant performance loss.
Ruben Ohana, Michael McCabe, Lucas Meyer +23 authors
A large-scale dataset collection, The Well, provides diverse numerical simulations for benchmarking machine learning models in physical systems simulation.
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