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。
Yuwei Guo, Ceyuan Yang, Anyi Rao +4 authors
A framework inserts a motion model into existing text-to-image models to enable animation, using video clips to learn motion priors while preserving model domain and diversity.
50 篇论文 · 按点赞排序
Nataniel Ruiz, Yuanzhen Li, Varun Jampani +6 authors
HyperDreamBooth leverages hypernetworks for efficient personalized face generation, significantly speeding up the process and reducing model size compared to existing methods.
Mostafa Dehghani, Basil Mustafa, Josip Djolonga +12 authors
NaViT, a flexible Vision Transformer, processes images of arbitrary resolutions and aspect ratios, improving training efficiency and performance across various computer vision tasks.
Subin Kim, Kyungmin Lee, June Suk Choi +3 authors
A novel method, Collaborative Score Distillation (CSD), based on Stein Variational Gradient Descent (SVGD), enhances consistency in text-to-image diffusion models across multiple images such as panoramas, videos, and 3D scenes.
Rui Zheng, Shihan Dou, Songyang Gao +24 authors
This report examines Reinfocement Learning with Human Feedback (RLHF) and proposes PPO-max to improve the stability of policy model training compared to other SFT models and ChatGPT.
Tao Ge, Jing Hu, Xun Wang +2 authors
The In-context Autoencoder (ICAE) compresses long contexts into memory slots for efficient LLM inference using a combination of pretraining and fine-tuning.
Xiangpeng Wei, Haoran Wei, Huan Lin +15 authors
PolyLM, a multilingual LLM trained on 640 billion tokens, enhances multilingual capabilities through bilingual data and curriculum learning, outperforming other models on multilingual tasks while maintaining English performance.
Waseem AlShikh, Manhal Daaboul, Kirk Goddard +4 authors
The Instruction Following Score (IFS) is introduced to evaluate and control instruction-following ability in language models, aiding in distinguishing between base and instruct models as well as serving as an early stopping criterion in training.
Yi Wang, Yinan He, Yizhuo Li +10 authors
InternVid, a large-scale video-text multimodal dataset, and ViCLIP, a video-text representation model based on ViT-L, enable advanced video and text understanding and generation tasks.
Vladislav Lialin, Namrata Shivagunde, Sherin Muckatira +1 authors
ReLoRA, a low-rank training method, demonstrates comparable performance to traditional training for large transformer language models and becomes more efficient with increased model size.
Aditya Sanghi, Pradeep Kumar Jayaraman, Arianna Rampini +4 authors
A pre-trained vision model's features enable generation of 3D shapes from sketches without paired datasets by leveraging synthetic renderings during training.
Quan Sun, Qiying Yu, Yufeng Cui +7 authors
Emu, a Transformer-based multimodal model, generates images and texts in various contexts and demonstrates superior performance across zero-shot and few-shot tasks compared to existing models.
Feng Li, Hao Zhang, Peize Sun +6 authors
Semantic-SAM, a universal image segmentation model, achieves semantic-awareness and multi-granularity by introducing decoupled classification and multi-choice learning, improving performance across various segmentation tasks.
Hugo Flores Garcia, Prem Seetharaman, Rithesh Kumar +1 authors
VampNet, a non-autoregressive transformer-based model, synthesizes and manipulates music with high-fidelity using variable masking and prompting techniques.
Zhenhailong Wang, Shaoguang Mao, Wenshan Wu +3 authors
Solo Performance Prompting (SPP) improves LLM performance by enabling internal persona collaboration, enhancing knowledge acquisition, reducing hallucination, and maintaining reasoning capabilities in complex tasks.
Nayoung Lee, Kartik Sreenivasan, Jason D. Lee +2 authors
Small transformers trained with chain-of-thought data demonstrate improved arithmetic learning, sample complexity, and convergence speed, highlighting the importance of high-quality, instructive training data.
Beibin Li, Konstantina Mellou, Bo Zhang +2 authors
A framework leveraging Large Language Models to provide interpretable insights into supply chain optimization outcomes without sharing proprietary data.
Grégoire Mialon, Quentin Garrido, Hannah Lawrence +3 authors
Joint embedding methods for self-supervised learning are applied to learn general-purpose representations of PDEs from heterogeneous data, outperforming baselines and enhancing neural solvers.
Suvir Mirchandani, Fei Xia, Pete Florence +6 authors
Large language models demonstrate zero-shot sequence completion and can extrapolate actions in robotics tasks, offering a potential transfer of word patterns to control policies.
Krishan Rana, Jesse Haviland, Sourav Garg +3 authors
SayPlan combines LLMs with 3D scene graphs to enable scalable and iterative task planning for robots in complex, large-scale environments.
Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski +3 authors
A 3D autodecoder framework generates high-quality static and articulated 3D assets from 2D images or monocular videos, using latent space embeddings, normalization, and 3D diffusion.
Tom Monnier, Jake Austin, Angjoo Kanazawa +2 authors
The approach generates a compact 3D world representation using textured primitives, optimized through differentiable rendering, and performs well compared to existing methods, supporting scene editing and simulations.
Shilong Zhang, Peize Sun, Shoufa Chen +5 authors
GPT4RoI enhances vision-language multimodal understanding by enabling region-level alignment through instruction tuning, offering better control, capacity, and composition capabilities.
Siddhartha Jain, Xiaofei Ma, Anoop Deoras +1 authors
A generalized self-consistency framework enhances the quality and consistency of large-scale language model outputs across various tasks with minimal computational overhead.
Moab Arar, Rinon Gal, Yuval Atzmon +4 authors
A domain-agnostic text-to-image personalization method uses contrastive regularization to improve semantic token representations, achieving superior performance and flexibility.
Shraman Pramanick, Yale Song, Sayan Nag +5 authors
EgoVLPv2 improves video-language pre-training by integrating cross-modal fusion directly into the backbones, enhancing representation and reducing fine-tuning costs across various tasks.
Yingqing He, Menghan Xia, Haoxin Chen +8 authors
A framework utilizes existing video clips to generate coherent storytelling videos by combining motion structure retrieval and text-guided video synthesis, offering flexibility and consistency.
Yu Gu, Sheng Zhang, Naoto Usuyama +8 authors
Distillation of large language models into task-specific student models via self-supervised learning enhances biomedical knowledge extraction tasks with cost and efficiency benefits.
Daoan Zhang, Weitong Zhang, Bing He +3 authors
DNAGPT is a generalized pre-trained model for DNA analysis that excels in various tasks through enhanced GPT architecture and diverse training objectives.
Yihan Cao, Yanbin Kang, Lichao Sun
InstructMining, a linear rule using natural language indicators, effectively selects high-quality instruction-following data for enhancing language model finetuning 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