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。
Ruoxi Sun, Sercan O Arik, Hootan Nakhost +4 authors
SQL-PaLM, an LLM-based Text-to-SQL model, outperforms existing systems in few-shot learning and fine-tuning settings and demonstrates superior generalization on diverse variants of the Spider dataset.
50 篇论文 · 按点赞排序
Yanyu Li, Huan Wang, Qing Jin +6 authors
Text-to-image diffusion models are optimized for mobile devices through efficient architecture and improved distillation techniques, achieving high-quality image generation with reduced computational requirements.
Chunyuan Li, Cliff Wong, Sheng Zhang +6 authors
A cost-efficient approach using a biomedical figure-caption dataset and GPT-4-generated instruction-following data trains a vision-language conversational assistant for biomedical images, achieving state-of-the-art performance on visual question answering tasks.
Pablo Pernias, Dominic Rampas, Marc Aubreville
Wuerstchen employs latent diffusion strategies for efficient text-to-image synthesis with reduced computational requirements and improved inference speed.
Yuxin Wen, John Kirchenbauer, Jonas Geiping +1 authors
A novel technique called Tree-Ring Watermarking robustly fingerprints diffusion model outputs by embedding a pattern into the initial noise vector and is invariant to various image transformations.
Nouha Dziri, Ximing Lu, Melanie Sclar +13 authors
Transformers solve compositional tasks by reducing multi-step reasoning to linearized subgraph matching, leading to performance decay with increased complexity.
Zeyue Xue, Guanglu Song, Qiushan Guo +4 authors
A text-conditional image diffusion model named RAPHAEL generates high-quality and artistic images by stacking mixture-of-experts layers, outperforming existing models in style switching and zero-shot FID score.
Weiyu Li, Xuelin Chen, Peizhuo Li +2 authors
GenMM efficiently generates diverse, high-quality motions from limited examples using generative motion matching with bidirectional visual similarity, demonstrating versatility in various scenarios beyond standard motion matching.
Jing Yu Koh, Daniel Fried, Ruslan Salakhutdinov
A method to integrate text-only large language models with image encoders and decoders through embedding space mapping achieves multimodal capabilities including image retrieval, generation, and dialogue.
Omri Avrahami, Kfir Aberman, Ohad Fried +2 authors
The method decomposes a single image into distinct text tokens for each concept and fine-tunes textual embeddings using masked diffusion loss to enable synthesis with fine-grained control.
Kihyuk Sohn, Nataniel Ruiz, Kimin Lee +11 authors
StyleDrop fine-tunes pre-trained text-to-image models to synthesize images in a specific style using minimal parameters and feedback.
Joseph Zhu, Peiye Zhuang
Reformulating optimization loss with diffusion prior and introducing depth supervision improve text-to-3D synthesis quality and multi-view consistency.
Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao +3 authors
Model dementia, caused by the use of model-generated content in training, affects the quality of generative models like LLMs, VAEs, and GMMs, highlighting the need for genuine human interaction data.
Maxwell Horton, Sachin Mehta, Ali Farhadi +1 authors
ByteFormer performs classification directly on raw file bytes, supporting multiple modalities and privacy-preserving inference without significant accuracy loss.
Yizhi Li, Ruibin Yuan, Ge Zhang +15 authors
MERT, a self-supervised learning model for music audio, uses a combination of RVQ-VAE and CQT-based teacher models to achieve state-of-the-art performance across 14 music understanding tasks.
Jinbo Xing, Menghan Xia, Yuxin Liu +9 authors
Make-Your-Video generates high-quality videos by combining text descriptions with motion structure using a Latent Diffusion Model with temporal modules, achieving better performance and temporal coherence than existing methods.
Yuma Koizumi, Heiga Zen, Shigeki Karita +7 authors
A new speech dataset called LibriTTS-R, derived by applying speech restoration to LibriTTS, improves sound quality and enhances neural end-to-end TTS performance.
Yuchao Gu, Xintao Wang, Jay Zhangjie Wu +10 authors
Mix-of-Show addresses challenges in decentralized multi-concept customization for text-to-image diffusion models using embedding-decomposed LoRA, gradient fusion, and regionally controllable sampling.
Paul S. Scotti, Atmadeep Banerjee, Jimmie Goode +9 authors
MindEye, a novel fMRI-to-image approach, retrieves and reconstructs viewed images with high accuracy using specialized submodules and diffusion prior in latent space.
Hila Chefer, Oran Lang, Mor Geva +5 authors
A method for decomposing text prompts into interpretable elements within text-to-image diffusion models reveals underlying structures and enables applications like bias detection and semantic manipulation.
Max Schwarzer, Johan Obando-Ceron, Aaron Courville +3 authors
The BBF RL agent achieves super-human performance in Atari 100K by scaling neural networks for value estimation with sample efficiency.
Xi Chen, Josip Djolonga, Piotr Padlewski +40 authors
PaLI-X, a multilingual vision and language model, achieves state-of-the-art performance on various tasks by scaling its components and task diversity.
Chi Zhang, Yiwen Chen, Yijun Fu +7 authors
A novel method generates high-quality, stylized 3D avatars using pre-trained image-text diffusion models and a GAN-based 3D generation network, addressing challenges in 3D data generation.
Yuechen Zhang, Jinbo Xing, Eric Lo +1 authors
RIVAL uses diffusion models and step-wise latent alignment to generate high-quality real-world image variations from a single exemplar, surpassing existing methods.
Rui Yang, Lin Song, Yanwei Li +4 authors
GPT4Tools enables efficient tool usage by open-source LLMs through LoRA optimization, improving accuracy and zero-shot capacity for unseen tools.
Taicheng Guo, Kehan Guo, Bozhao nan +5 authors
Evaluation of GPT models on a comprehensive benchmark of practical chemistry tasks reveals varying performance across different types of chemical challenges.
Yuan Gong, Youxin Pang, Xiaodong Cun +7 authors
A system for generic interactive story visualization uses interconnected components to handle multiple characters, editable layouts, and detailed image generation, enhancing flexibility and quality across diverse stories.
Wen Yang, Chong Li, Jiajun Zhang +1 authors
BigTrans, an enhanced version of LLaMA, achieves competitive multilingual translation performance across over 100 languages.
Yuandong Tian, Yiping Wang, Beidi Chen +1 authors
The analysis of SGD training dynamics in a one-layer transformer reveals that self-attention operates as a discriminative scanning algorithm, focusing on distinct tokens for prediction and demonstrating a phase transition controlled by layer learning rates.
Murray Shanahan, Kyle McDonell, Laria Reynolds
Dialogue agents' behaviors, such as deception and self-awareness, are analyzed through role-play to avoid anthropomorphism.
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