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
Zehui Chen, Kuikun Liu, Qiuchen Wang +5 authors
Agent-FLAN fine-tunes LLMs for agent tasks by reorganizing the training corpus, reducing hallucinations, and enhancing agent capability without degrading general performance.
Shanchuan Lin, Xiao Yang
AnimateDiff-Lightning uses progressive adversarial diffusion distillation for efficient video generation, enhancing style compatibility through joint distillation of multiple diffusion models.
Ruyi Xu, Yuan Yao, Zonghao Guo +7 authors
LLaVA-UHD, a large multimodal model, efficiently processes images of any aspect ratio and high resolution by modularizing and compressing image tokens, outperforming established models on various benchmarks.
Peter Kocsis, Julien Philip, Kalyan Sunkavalli +2 authors
LightIt, a method for controllable illumination in image generation, uses shading and normal maps to produce high-quality images with consistent lighting, achieving results comparable to specialized relighting techniques.
Ge Zhu, Juan-Pablo Caceres, Zhiyao Duan +1 authors
MusicHiFi is an efficient high-fidelity stereophonic vocoder using a cascade of GANs to improve audio quality and spatialization control.
Yuval Alaluf, Elad Richardson, Sergey Tulyakov +2 authors
External concept heads augment VLMs to integrate user-specific concepts into image captioning and question-answering, demonstrating generalization to unseen images.
Paul S. Scotti, Mihir Tripathy, Cesar Kadir Torrico Villanueva +8 authors
High-quality visual perception reconstructions from brain activity are achieved using minimal fMRI data through a novel functional alignment and cross-subject fine-tuning approach, leveraging CLIP and Stable Diffusion XL.
Yinghao Xu, Zifan Shi, Wang Yifan +5 authors
GRM, a transformer-based model, efficiently reconstructs 3D assets from sparse images using pixel-aligned Gaussians and achieves superior quality and speed compared to alternatives.
Hadi Alzayer, Zhihao Xia, Xuaner Zhang +3 authors
A generative model utilizes video-based supervision to synthesize photorealistic edits from coarse layouts, transferring details and harmonizing lighting and interactions.
Vidhi Jain, Maria Attarian, Nikhil J Joshi +10 authors
Vid2Robot is a framework that uses video input to infer and execute robotic manipulation tasks by leveraging cross-attention and contrastive losses for improved policy performance and emergent capabilities.
Hansheng Chen, Ruoxi Shi, Yulin Liu +5 authors
MVEdit uses ancestral sampling to denoise multi-view images and generate 3D textured meshes, achieving high-quality 3D consistency through a 3D Adapter and surpassing score distillation in both speed and quality.
Arthur Douillard, Qixuan Feng, Andrei A. Rusu +7 authors
A modular architecture and training approach, DiPaCo, improves large-scale neural network training by distributing computation and optimizing communication across heterogeneous workers, achieving performance exceeding a billion-parameter dense transformer.
Antoine Guédon, Vincent Lepetit
Gaussian Frosting, an adaptive mesh-based Gaussian representation, enhances real-time rendering and editing of complex 3D effects by combining a base mesh with a variable thickness Gaussian layer to capture fine details and volumetric surfaces.
Olga Golovneva, Zeyuan Allen-Zhu, Jason Weston +1 authors
Reverse training enhances LLM performance on reversal tasks by doubling the training data through forward and reverse direction training while preserving entities.
Yue Fan, Xiaojian Ma, Rujie Wu +4 authors
A novel unified memory mechanism in VideoAgent enhances video understanding by combining foundation models, improving performance on long-horizon video benchmarks.
Haiwen Feng, Zheng Ding, Zhihao Xia +4 authors
Time Reversal Fusion enables video generation by fusing forward and backward denoising paths conditioned on start and end frames, outperforming existing methods in generating smooth, complex motions, and 3D-consistent views.
Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang +5 authors
MOTIA, a diffusion-based pipeline, achieves video outpainting with high quality and flexibility by leveraging input-specific adaptation and pattern-aware outpainting phases.
Quankai Gao, Qiangeng Xu, Zhe Cao +5 authors
Gaussian flow connects 3D Gaussian dynamics and pixel velocities, enabling effective 4D content generation and novel view synthesis with improved Gaussian dynamics.
Victor Carbune, Hassan Mansoor, Fangyu Liu +4 authors
A method transfers reasoning capabilities from large-language models to vision-language models, achieving state-of-the-art performance on ChartQA and superior performance on PlotQA and FigureQA by enhancing chart representation and training with a larger, synthesized dataset and multitask loss.
Yushi Lan, Fangzhou Hong, Shuai Yang +5 authors
A novel 3D diffusion framework, LN3Diff, uses a 3D-aware architecture and VAE to enable fast, high-quality 3D generation, outperforming existing methods in reconstruction speed and quality on various datasets.
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