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
Yael Vinker, Tamar Rott Shaham, Kristine Zheng +3 authors
SketchAgent, a language-driven method, generates and refines sketches through dynamic conversations using off-the-shelf LLMs and string-based actions processed into vector graphics.
Zechen Bai, Jianxiong Gao, Ziteng Gao +4 authors
Factorized Quantization (FQ) improves image generation by decomposing large codebooks into smaller, diverse sub-codebooks, enhancing scalability and reconstruction quality through disentanglement regularization and semantic representation learning.
Kaizhao Liang, Lizhang Chen, Bo Liu +1 authors
A single-line modification to momentum-based optimizers preserves Adam's convergence properties and achieves significant speed-up in transformer pretraining.
Zun Wang, Jialu Li, Han Lin +2 authors
DreamRunner integrates scene structuring, motion retrieval, and spatial-temporal attention to generate coherent, motion-rich storytelling videos from text scripts.
Davide Paglieri, Bartłomiej Cupiał, Samuel Coward +10 authors
BALROG, a novel benchmark, evaluates the agentic capabilities of large language models and vision language models across diverse, challenging games with fine-grained metrics, revealing significant performance gaps, especially in vision-based decision-making.
Shengqu Cai, Eric Chan, Yunzhi Zhang +3 authors
Diffusion Self-Distillation uses a pre-trained text-to-image model to generate a dataset for fine-tuning into a text+image-to-image model, improving identity-preservation generation without test-time optimization.
Xin Yu, Ze Yuan, Yuan-Chen Guo +6 authors
A large diffusion model trained directly in UV texture space generates high-resolution texture maps from text prompts and images, supporting applications like inpainting, completion, and synthesis.
Gyeongjin Kang, Jisang Yoo, Jihyeon Park +5 authors
SelfSplat integrates 3D Gaussian Splatting with self-supervised depth and pose estimation to achieve high-quality, pose-free, 3D reconstruction from unposed multi-view images, demonstrating superior performance and generalization.
Jan Held, Renaud Vandeghen, Abdullah Hamdi +6 authors
3D Convex Splatting achieves superior scene reconstruction and novel view synthesis by using smooth convex shapes, offering improved performance and efficiency compared to 3D Gaussian Splatting.
Zhiyang Guo, Jinxu Xiang, Kai Ma +3 authors
Make-It-Animatable is a data-driven method that converts 3D humanoid models into animatable avatars quickly and effectively using a particle-based shape autoencoder and a coarse-to-fine strategy.
Bencheng Liao, Shaoyu Chen, Haoran Yin +8 authors
A truncated diffusion policy with a cascade decoder improves real-time robustness and diversity in robotic policy learning for autonomous driving.
Carlo Alberto Barbano, Luca Molinaro, Emanuele Aiello +1 authors
Knowledge Transfer uses textual descriptions to efficiently introduce and enhance novel concepts in multimodal models, improving zero-shot performance.
Wang Bill Zhu, Deqing Fu, Kai Sun +8 authors
VisualLens extracts, filters, and refines image representations from users' visual histories to enhance personalization, outperforming existing methods and GPT-4o in recommendation tasks.
Yiheng Li, Ruibing Hou, Hong Chang +2 authors
UniPose, leveraging Large Language Models and unified visual encoders, comprehends, generates, and edits human poses across multiple modalities with superior performance.
Ahmed Heakl, Chaimaa Abi, Rania Hossam +1 authors
A lightweight LLM-based transpiler, CRT, converts x86 assembly to ARM assembly, bridging architectural gaps with high accuracy and improved performance on ARM hardware.
Songhao Han, Wei Huang, Hairong Shi +7 authors
VideoEspresso, a novel dataset with high-quality VideoQA pairs and multimodal annotations, enhances LVLMs for video reasoning through a semantic-aware pipeline and hybrid collaboration framework.
Jaemin Kim, Bryan S Kim, Jong Chul Ye
Free$^2$Guide is a gradient-free framework that enhances text alignment in text-to-video generation using path integral control and non-differentiable reward functions, integrating large Vision-Language Models.
Jeovane Honorio Alves, Radu State, Cinthia Obladen de Almendra Freitas +1 authors
LongKey, a novel framework using an encoder-based language model and max-pooling embedder, extracts keyphrases from lengthy documents, outperforming existing methods on comprehensive datasets.
Emanuele Aiello, Umberto Michieli, Diego Valsesia +2 authors
DreamCache is a scalable and efficient personalized image generation method that leverages cached reference image features and conditioning adapters for high-quality and text-aligned image creation with reduced parameter overhead.
Zigeng Chen, Xinyin Ma, Gongfan Fang +1 authors
Collaborative Decoding (CoDe) enhances Visual Auto-Regressive (VAR) modeling efficiency by integrating a large and small model to reduce memory usage and computation time without significantly affecting image quality.
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