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。
Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho +3 authors
A study on Self-Rewarding Language Models shows that using LLM-as-a-Judge prompting for iterative DPO training enhances both instruction-following and self-reward generation, leading to superior performance compared to existing systems.
31 篇论文 · 按点赞排序
Lianghui Zhu, Bencheng Liao, Qian Zhang +3 authors
A new vision backbone using bidirectional Mamba blocks (Vim) outperforms established vision transformers like DeiT in terms of performance and efficiency on high-resolution image tasks.
Qixun Wang, Xu Bai, Haofan Wang +2 authors
InstantID is a diffusion model-based solution for personalized image synthesis that uses a single facial image and integrates with pre-trained models while maintaining high face fidelity and efficiency.
Yue Liu, Yunjie Tian, Yuzhong Zhao +5 authors
VMamba, a novel state space model architecture, combines global receptive fields and dynamic weights from ViTs with linear complexity, outperforming established models as image resolution increases.
Haoran Xu, Amr Sharaf, Yunmo Chen +5 authors
Contrastive Preference Optimization (CPO) enhances moderate-sized LLMs in machine translation, enabling them to match or exceed state-of-the-art model performance.
Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai +5 authors
Autoregressive pre-training for vision models (AIM) scales similarly to LLMs, showing improved performance with more data and parameters, and does not exhibit performance saturation.
Zihan Liu, Wei Ping, Rajarshi Roy +3 authors
ChatQA uses a two-stage instruction tuning method with a dense retriever to improve zero-shot conversational QA performance, outperforming GPT-4 on multiple datasets.
Trung Quoc Luong, Xinbo Zhang, Zhanming Jie +3 authors
Reinforced Fine-Tuning (ReFT) improves the generalizability of large language models in reasoning tasks like math problem-solving by using reinforcement learning to learn from multiple reasoning paths.
Jie Qin, Jie Wu, Weifeng Chen +6 authors
DiffusionGPT combines Large Language Models with diffusion models and Trees-of-Thought to handle diverse prompts and integrate domain-expert models, demonstrating superior image synthesis performance across various domains.
Jiaheng Liu, Zhiqi Bai, Yuanxing Zhang +11 authors
E 2 -LLM extends long-context capabilities in LLMs with reduced computation and data requirements through single training, short context windows, and novel position embedding augmentations.
Alisa Liu, Xiaochuang Han, Yizhong Wang +3 authors
Proxy-tuning, a lightweight decoding-time algorithm, customizes large language models efficiently using smaller LMs without accessing their weights, achieving competitive performance and better factual truthfulness.
Chung Min Kim, Mingxuan Wu, Justin Kerr +3 authors
GARField extracts semantically meaningful hierarchical groupings from 3D scenes using scale-conditioned 3D affinity feature fields derived from 2D masks.
Baoxiong Jia, Yixin Chen, Huangyue Yu +5 authors
A large-scale 3D vision-language dataset and pre-training framework achieve state-of-the-art results in 3D visual grounding tasks.
Ioana Bica, Anastasija Ilić, Matthias Bauer +8 authors
SPARC pretrains fine-grained multimodal representations using sparse contrastive alignment, improving both global and local information in vision-language models and enhancing performance on classification and region-level tasks.
Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit +3 authors
The paper identifies and critiques limitations of FID as an evaluation metric for image generation, proposing a new metric, CMMD, based on CLIP embeddings and maximum mean discrepancy for more reliable evaluations.
Xiaofeng Wang, Zheng Zhu, Guan Huang +3 authors
WorldDreamer, a world model inspired by large language models, excels at unsupervised visual sequence modeling, generating videos in various scenarios with tasks such as text-to-video conversion and image-to-video synthesis.
Ludan Ruan, Lei Tian, Chuanwei Huang +2 authors
The Unified-modal Video Generation system uses Multi-condition Cross Attention and Biased Gaussian Noise to handle diverse video generation tasks across text and image modalities, achieving superior performance on benchmarks.
Yikai Zhang, Junlong Li, Pengfei Liu
A novel extension to rotary position embedding, adjusted by modifying the base frequency and scaling attention logits, enhances LLMs' context window efficiently and robustly across various tasks.
Connor Holmes, Masahiro Tanaka, Michael Wyatt +8 authors
DeepSpeed-FastGen enhances the deployment of large language models by introducing Dynamic SplitFuse, improving throughput and latency compared to existing frameworks.
Jay Zhangjie Wu, Guian Fang, Haoning Wu +11 authors
A new metric, T2VScore, is introduced to evaluate text-to-video generation by combining text-video alignment and video quality, addressing the limitations of current automated metrics.
Andreas Engelhardt, Amit Raj, Mark Boss +8 authors
SHINOBI reconstructs 3D shapes, materials, and illumination from unconstrained images using implicit shape representation and joint optimization of BRDF and camera alignment.
Haoxin Chen, Yong Zhang, Xiaodong Cun +4 authors
A method is proposed to generate high-quality videos using low-quality videos and synthesized images by fine-tuning spatial modules of video models extended from Stable Diffusion.
Michael Kölle, Gerhard Stenzel, Jonas Stein +3 authors
Quantum diffusion models address sampling speed and parameter requirements in classical diffusion models by using variational quantum circuits, demonstrating superior performance and one-step image generation.
Tan Dat Nguyen, Ji-Hoon Kim, Youngjoon Jang +2 authors
FreGrad is a lightweight and fast diffusion-based vocoder that uses discrete wavelet transform for simplified feature representation, frequency-aware dilated convolution for improved frequency accuracy, and optimization techniques to achieve faster training and inference while maintaining audio quality.
Nanye Ma, Mark Goldstein, Michael S. Albergo +3 authors
Scalable Interpolant Transformers (SiT), built on Diffusion Transformers (DiT), enhance generative models through a flexible interpolant framework, achieving superior performance on ImageNet 256x256 with tunable diffusion coefficients.
Bo Liu, Rachita Chhaparia, Arthur Douillard +5 authors
Asynchronous Local-SGD with delayed Nesterov momentum and adaptive local training steps improves performance and clock time for large language models training.
Yu-Ying Yeh, Jia-Bin Huang, Changil Kim +8 authors
TextureDreamer uses personalized geometry-aware score distillation to transfer detailed textures from a few images to arbitrary 3D shapes, surpassing existing methods.
Antoine Mercier, Ramin Nakhli, Mahesh Reddy +4 authors
HexaGen3D fine-tunes a 2D diffusion model to generate high-quality 3D assets from textual prompts using 6 orthographic projections and a latent triplane, offering better quality-to-latency trade-offs than existing methods.
Jonghyun Lee, Hansam Cho, Youngjoon Yoo +2 authors
A conditional diffusion model with depth disentanglement and soft guidance improves localization and semantic composition of objects in generated images.
Zhao Wang, Aoxue Li, Enze Xie +4 authors
A novel framework CustomVideo enhances multi-subject text-to-video generation by using an attention control strategy in latent space and object segmentation masks to produce identity-preserving videos.
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