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
Qinghe Wang, Xu Jia, Xiaomin Li +4 authors
StableIdentity injects learned identity from a single face image into various contexts, including video and 3D, ensuring identity consistency and flexible edits.
Chen Bai, Zeman Shao, Guoxiang Zhang +11 authors
A novel framework for realistic video simulation, AnyObject in Any Scene, achieves high geometric, lighting, and photorealism by integrating objects, simulating lighting and shadows, and refining video output through style transfer.
Tokiniaina Raharison Ralambomihanta, Shahrad Mohammadzadeh, Mohammad Sami Nur Islam +2 authors
Knowledge distillation using the Hyena mechanism enhances both accuracy and efficiency in LLM pre-training, replacing attention heads and addressing long contextual information processing.
Kehang Han, Kathleen Kenealy, Aditya Barua +2 authors
Text diffusion models show promise in code synthesis and extractive QA, outperforming autoregressive models in some cases and offering faster long text generation through lightweight adaptation.
Edward Bartrum, Thu Nguyen-Phuoc, Chris Xie +5 authors
A novel text-guided 3D scene editing model replaces specific objects while maintaining 3D consistency across multiple viewpoints.
Xuandong Zhao, Xianjun Yang, Tianyu Pang +4 authors
A weak-to-strong jailbreaking attack uses smaller, less aligned language models to exploit larger, aligned models, highlighting a significant security concern in language model alignment.
Guihong Li, Hsiang Hsu, Chun-Fu +2 authors
A novel framework and algorithm for machine unlearning in image-to-image generative models is introduced, showing minimal performance loss and compliance with data retention policies.
Marcos V. Conde, Gregor Geigle, Radu Timofte
InstructIR is an all-in-one image restoration method that leverages human-written instructions to guide the restoration process, achieving state-of-the-art results across multiple tasks.
Yifan Peng, Jinchuan Tian, William Chen +9 authors
E-Branchformer-based OWSM v3.1 models improve speech processing efficiency and performance without additional training data.
Yikai Wang, Chenjie Cao, Qiaole Dong +2 authors
A diffusion generative model, using SEELE framework, handles dynamic subject repositioning by reformulating the task into unified prompt-guided inpainting with task inversion techniques.
Xiaojun Wu, Dixiang Zhang, Ruyi Gan +6 authors
Taiyi-Diffusion-XL is a bilingual text-to-image model that extends CLIP and Stable-Diffusion-XL for Chinese and English, improving vocabulary and vision-language capabilities, and achieving superior bilingual image generation.
Zihao Wang, Chirag Nagpal, Jonathan Berant +4 authors
The study examines the transformation and aggregation of reward models to align language models with human preferences, identifying a method that improves poor outputs and aggregates rewards logically.
Jianxiang Lu, Cong Xie, Hui Guo
A novel fine-tuning method using prototypes and class-characterizing regularization improves object-driven generalization and fidelity in text-to-image generation models.
Guocheng Qian, Junli Cao, Aliaksandr Siarohin +12 authors
Amortized Text-to-Mesh (AToM) generates high-quality textured meshes efficiently from text prompts using a novel triplane-based architecture and achieves superior accuracy and generalization compared to existing methods.
Zhenyu Wang, Enze Xie, Aoxue Li +3 authors
CompAgent, a training-free method using an LLM agent, improves compositional text-to-image generation by decomposing complex text prompts, reasoning, and integrating multiple tools to enhance image quality and controllability.
Elias Stengel-Eskin, Archiki Prasad, Mohit Bansal
Refactoring for Generalizable Abstraction Learning (ReGAL) enhances program synthesis by creating a library of reusable functions, improving the accuracy of large language models across diverse tasks.
Jingyu Zhuang, Di Kang, Yan-Pei Cao +3 authors
TIPEditor, a text- and image-driven 3D scene editing framework, uses 2D personalization and 3D Gaussian splatting to achieve accurate and aligned edits within specified bounding boxes.
Jiezhi Yang, Khushi Desai, Charles Packer +4 authors
A method for predicting future 3D scenes using a conditional auto-encoded radiance field, mapping images to plausible 3D latent configurations and leveraging probabilistic and auto-regressive predictions for autonomous driving applications.
Xiaoran Fan, Tao Ji, Changhao Jiang +21 authors
The ensemble experts technique with a fusion network improves VLM performance by combining visual encoders and optimizing positional encoding for lengthy image feature sequences.
Yuhang Zang, Hanlin Goh, Josh Susskind +1 authors
A new method, OGEN, improves out-of-distribution generalization in finetuned vision-language models by synthesizing features for unknown classes and using adaptive self-distillation.
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