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
Yueming Pan, Ruoyu Feng, Qi Dai +5 authors
Semantic-First Diffusion (SFD) enhances image generation by asynchronously denoising semantic and texture latents, improving convergence and quality.
Yunhong Lu, Yanhong Zeng, Haobo Li +9 authors
The paper introduces Reward Forcing, which enhances video generation by updating sink tokens with EMA-Sink and using Rewarded Distribution Matching Distillation to prioritize dynamic content.
Siyuan Yang, Yang Zhang, Haoran He +4 authors
TACO, a test-time-scaling framework with a pseudo-count estimator, enhances the inference stability and success rates of Vision-Language-Action models in downstream tasks by preventing distribution shifts.
Zeren Chen, Xiaoya Lu, Zhijie Zheng +6 authors
Geometrically-Constrained Agent (GCA) addresses the semantic-to-geometric gap in vision language models by decoupling semantic analysis and task solving with formal constraints, achieving state-of-the-art performance in spatial reasoning.
Haochen Tian, Tianyu Li, Haochen Liu +11 authors
A simulation framework improves autonomous driving by generating diverse, high-fidelity driving scenarios, leading to better generalization and robustness in real-world testing.
Yongsheng Yu, Wei Xiong, Weili Nie +3 authors
PixelDiT is a single-stage, end-to-end diffusion model that operates directly in pixel space, overcoming the limitations of latent-space modeling by using a dual-level transformer architecture and achieving competitive performance in image and text-to-image generation.
Ziheng Ouyang, Yiren Song, Yaoli Liu +4 authors
ImageCritic addresses detail inconsistency in image generation through reference-guided post-editing, using attention alignment loss and a detail encoder.
Kairun Wen, Yuzhi Huang, Runyu Chen +16 authors
DynamicVerse is a framework that models dynamic real-world videos by integrating large vision, geometric, and multimodal models to produce a comprehensive 4D multimodal dataset, achieving superior performance in video depth estimation, camera pose estimation, and camera intrinsics estimation.
Yonggan Fu, Xin Dong, Shizhe Diao +12 authors
The study identifies key architectural factors and efficient operators to optimize small language models for real-device latency, introducing the Nemotron-Flash family for improved accuracy and efficiency.
Jintian Zhang, Kewei Xu, Jingsheng Zheng +10 authors
InnoGym is a benchmark and framework that evaluates the innovation potential of AI agents using performance gain and novelty metrics, highlighting a gap between creativity and effectiveness.
Dongyang Liu, Peng Gao, David Liu +8 authors
The study reveals that in text-to-image generation, CFG Augmentation is the primary driver of few-step distillation in Distribution Matching Distillation (DMD), while the distribution matching term acts as a regularizer.
Kaituo Feng, Manyuan Zhang, Hongyu Li +11 authors
OneThinker, an all-in-one multimodal reasoning model, unifies image and video understanding across various tasks using RL and demonstrates strong performance and knowledge transfer.
Jiawei Ren, Yan Zhuang, Xiaokang Ye +20 authors
SimWorld, a new Unreal Engine 5-based simulator, enables the development and evaluation of LLM/VLM agents in realistic, real-world-like settings with diverse physical and social reasoning scenarios.
Jung Yi, Wooseok Jang, Paul Hyunbin Cho +3 authors
Deep Forcing, a training-free method, enhances real-time video diffusion by addressing temporal repetition and motion issues through Deep Sink and Participative Compression, achieving high-quality, long-duration video generation.
Zhennan Chen, Junwei Zhu, Xu Chen +6 authors
DiP, a pixel space diffusion framework, combines a Diffusion Transformer and a Patch Detailer Head to achieve computational efficiency and high-quality image generation without using VAEs.
Zhuobai Dong, Rui Zhao, Songjie Wu +5 authors
Using phase-aware LoRA adapters, diffusion models achieve efficient acceleration and strong generalization with minimal retraining.
Jiangtao Wu, Shihao Li, Zhaozhou Bian +7 authors
The ViDiC task and ViDiC-1K dataset evaluate Multimodal Large Language Models' ability to describe differences between video pairs, addressing limitations in capturing motion continuity and event evolution.
Shijia Yang, Yunong Liu, Bohan Zhai +5 authors
CaptionQA evaluates caption utility by measuring their effectiveness in supporting downstream tasks, revealing significant gaps compared to traditional image-QA benchmarks.
Jiaming Tang, Yufei Sun, Yilong Zhao +7 authors
VLASH is an asynchronous inference framework for Vision-Language-Action models that achieves high speed and low latency without sacrificing accuracy, enabling precise robotic tasks like ping-pong and whack-a-mole.
Guanjie Chen, Shirui Huang, Kai Liu +5 authors
Flash-DMD is a framework that combines efficient timestep-aware distillation and reinforcement learning to accelerate and stabilize the training of generative diffusion models, achieving high generation quality with reduced computational cost.
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