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
Senthil Purushwalkam, Akash Gokul, Shafiq Joty +1 authors
BootPIG architecture enhances text-to-image generation models with personalization capabilities, using reference images, to improve the fidelity to the desired appearance while maintaining alignment with textual prompts.
Boxiao Pan, Zhan Xu, Chun-Hao Paul Huang +4 authors
ActAnywhere, a generative model utilizing video diffusion, automates coherent background generation for videos, aligning with foreground motion and artistic conditions.
Wei Li, Xue Xu, Jiachen Liu +1 authors
A multimodal conditional diffusion framework named UNIMO-G generates images from both text and visual prompts, capable of high-fidelity synthesis of intricate details and diverse image entities.
Michael Ahn, Debidatta Dwibedi, Chelsea Finn +25 authors
AutoRT uses vision-language and large language models to enable operational robots to perform diverse tasks in unseen scenarios with minimal human supervision, enhancing data collection and alignment with human preferences.
Feng Lin, Hanling Yi, Hongbin Li +4 authors
BiTA, a parameter-efficient bi-directional tuning method for LLMs, accelerates inference via parallel draft candidate generation and verification without additional memory costs.
Boyi Li, Jathushan Rajasegaran, Yossi Gandelsman +2 authors
A diffusion model-based framework animates people from a single image using learned priors for unseen parts and a diffusion-based rendering pipeline controlled by 3D human poses.
Minglin Chen, Longguang Wang, Weihao Yuan +6 authors
A multi-view sketch-guided text-to-3D generation framework (Sketch2NeRF) uses pretrained 2D diffusion models and NeRF to produce high-fidelity 3D content with fine-grained sketch control, achieving state-of-the-art performance in sketch similarity and text alignment.
Peiqin Lin, Shaoxiong Ji, Jörg Tiedemann +2 authors
MaLA-500, a large language model covering 534 languages, achieves state-of-the-art in-context learning by extending the LLaMA 2 vocabulary with Glot500-c.
Ke Ye, Heinrich Jiang, Afshin Rostamizadeh +6 authors
SpacTor, a hybrid training procedure combining span corruption and token replacement detection with a two-stage curriculum, enhances pre-training efficiency and performance for large language models.
Rohan Wadhawan, Hritik Bansal, Kai-Wei Chang +1 authors
A new benchmark, ConTextual, evaluates LMMs' context-sensitive text-rich visual reasoning, revealing a significant performance gap compared to humans.
Zhichao Wang, Yuanzhe Chen, Xinsheng Wang +4 authors
StreamVoice is a real-time, zero-shot voice conversion model that uses a fully causal context-aware language model with strategies to handle missing context, enabling conversion without future look-ahead.
Ege Ozguroglu, Ruoshi Liu, Dídac Surís +4 authors
Pix2gestalt uses conditional diffusion models, trained on synthetic occluded-object datasets, to achieve zero-shot amodal segmentation better than supervised approaches and enhances object recognition and 3D reconstruction in occlusions.
Nisha Huang, Weiming Dong, Yuxin Zhang +5 authors
CreativeSynth, a diffusion-based framework with multimodal attention, enhances artistic image editing by accurate style and content manipulation while preserving aesthetic quality.
Peiqi Liu, Yaswanth Orru, Chris Paxton +2 authors
OK-Robot, a new framework combining Vision-Language Models, navigation primitives, and grasping primitives, achieves high success rates in open-ended pick-and-drop tasks across various home environments.
Xin Yuan, Jinoo Baek, Keyang Xu +2 authors
An efficient diffusion-based method inflates a text-to-image super-resolution framework to generate high-quality, temporally coherent text-to-video super-resolution content.
Haoyu Xiong, Russell Mendonca, Kenneth Shaw +1 authors
An adaptive learning framework and hardware platform enable a robot to perform mobile manipulation of various articulated objects in unstructured environments with high success rates through online adaptation.
Taeksoo Kim, Byungjun Kim, Shunsuke Saito +1 authors
GALA uses a pretrained 2D diffusion model and pose-guided Score Distillation Sampling to decompose single-layer 3D human meshes into multi-layered assets for use in novel avatars.
W. Ronny Huang, Cyril Allauzen, Tongzhou Chen +7 authors
A non-autoregressive LM-fused ASR system using USM and PaLM 2 improves WER across languages and highlights key factors impacting performance.
Susan Lin, Jeremy Warner, J. D. Zamfirescu-Pereira +8 authors
Rambler, an LLM-powered GUI, simplifies dictation post-processing by enabling gist-level text manipulation and macro revisions, outperforming traditional tools in facilitating user-controlled iterative composition.
Matthew Kowal, Achal Dave, Rares Ambrus +3 authors
The study introduces a method to identify and rank spatiotemporal concepts in video transformer models, enhancing interpretability and performance on fine-grained tasks.
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