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
Min Zhao, Rongzhen Wang, Fan Bao +2 authors
ControlVideo enhances text-driven video editing by integrating text-to-image diffusion models and ControlNet, improving video fidelity and temporal consistency while maintaining source structure.
Yonglong Tian, Lijie Fan, Phillip Isola +2 authors
Training self-supervised methods on synthetic images generated by a text-to-image model like Stable Diffusion can yield high-performing visual representations matching or exceeding those trained on real images.
Shaozhe Hao, Kai Han, Shihao Zhao +1 authors
ViCo, a plug-in method using diffusion models, achieves high-quality personalized text-to-image generation with minimal parameter training through an image attention module and attention-based object masks.
Bailin Wang, Zi Wang, Xuezhi Wang +3 authors
Grammar prompting enhances LLMs' performance on DSL generation tasks by incorporating domain-specific grammars expressed in BNF.
Jiawei Huang, Yi Ren, Rongjie Huang +7 authors
Make-an-Audio 2, a latent diffusion-based text-to-audio synthesis method, improves semantic alignment and temporal consistency through pre-trained language models, structured-text encoding, a Transformer-based diffusion denoiser, and data augmentation.
Tianjun Zhang, Yi Zhang, Vibhav Vineet +2 authors
Control-GPT enhances text-to-image generation by using GPT-4-generated TikZ sketches as guides, improving spatial arrangement and object positioning accuracy.
Sadhika Malladi, Tianyu Gao, Eshaan Nichani +4 authors
MeZO, a memory-efficient zeroth-order optimizer, fine-tunes large language models with significantly reduced memory usage while maintaining performance comparable to backpropagation.
Xizhou Zhu, Yuntao Chen, Hao Tian +10 authors
A framework integrating Large Language Models with text-based knowledge achieves superior performance in Minecraft by surpassing traditional RL methods and completing complex tasks.
Fu-Yun Wang, Wenshuo Chen, Guanglu Song +3 authors
A novel paradigm extends short video diffusion models to handle long, multi-segment, text-conditioned video generation and editing without additional training.
Jing Gu, Yilin Wang, Nanxuan Zhao +8 authors
Photoswap enables seamless subject swapping in images using pre-trained diffusion models with self-attention and cross-attention manipulation, preserving pose and coherence.
Bill Yuchen Lin, Yicheng Fu, Karina Yang +6 authors
SwiftSage, a dual-module agent framework combining behavior cloning and large language models, excels at complex interactive reasoning tasks by integrating intuitive and deliberate thinking.
Ruibo Liu, Ruixin Yang, Chenyan Jia +5 authors
A novel training paradigm enables language models to learn from simulated social interactions, improving their alignment with societal norms and values compared to existing methods.
Yixuan Su, Tian Lan, Huayang Li +3 authors
PandaGPT integrates multimodal encoders and large language models to perform complex tasks involving images, videos, and audio, demonstrating zero-shot cross-modal understanding.
Jikun Kang, Romain Laroche, Xindi Yuan +3 authors
An internal working memory module improves training efficiency and generalization for decision-making in Atari and meta-world tasks by addressing the forgetting phenomenon.
Mingchen Zhuge, Haozhe Liu, Francesco Faccio +23 authors
Natural language-based societies of minds (NLSOMs) consisting of diverse neural network experts improve multimodal zero-shot reasoning and can solve various practical AI tasks by communicating through a universal symbolic language.
Ying Fan, Olivia Watkins, Yuqing Du +7 authors
Using online reinforcement learning with policy gradient and KL regularization to fine-tune text-to-image diffusion models improves image-text alignment and image quality compared to supervised fine-tuning.
Tianle Cai, Xuezhi Wang, Tengyu Ma +2 authors
A proposed framework, LATM, allows large language models (LLMs) to create and use their own tools for problem-solving, enhancing efficiency and cost-effectiveness.
Ge Yuan, Xiaodong Cun, Yong Zhang +5 authors
A novel method personalizes pre-trained diffusion models using minimal data and parameters, enabling high-quality image generation of new identities and interactions.
Hao Liu, Pieter Abbeel
Blockwise Parallel Transformer (BPT) enhances memory efficiency and handles longer sequences by optimizing self-attention and feedforward network computations.
Gowthami Somepalli, Vasu Singla, Micah Goldblum +2 authors
Text-to-image diffusion models, such as Stable Diffusion, frequently replicate training data during inference, especially with text conditioning; proposed techniques address this by randomizing and augmenting image captions.
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