MagicMirror uses 340K annotated images and GRPO-trained VLM for fine-grained artifact assessment, advancing model defect detection.

Jia Wang, Jie Hu, Xiaoqi Ma et al.

The OnlineHOI framework achieves state-of-the-art results in online human-object interaction generation and perception on Core4D and OAKINK2 datasets.

Yihong Ji, Yunze Liu, Yiyao Zhuo et al.

InterAct enhances 3D human-object interaction data quality with a unified optimization framework, expanding the dataset to 30.70 hours.

Sirui Xu, Dongting Li, Yucheng Zhang et al.

Dark-ISP enhances low-light object detection by directly processing Bayer RAW images, outperforming existing methods.

Jiasheng Guo, Xin Gao, Yuxiang Yan et al.

RewardDance scales reward models up to 26B parameters, improving visual generation quality and robustness against reward hacking.

Jie Wu, Yu Gao, Zilyu Ye et al.

EVDI++ uses event cameras for video deblurring and interpolation, improving video quality.

Chi Zhang, Xiang Zhang, Chenxu Jiang et al.

VIRAL aligns visual representations to enhance multimodal large models' performance on visual tasks, with an average improvement of 9.4%.

Heeji Yoon, Jaewoo Jung, Junwan Kim et al.

UFC method achieves spatial control in diffusion models with few samples, nearing fully supervised baselines.

Kiet T. Nguyen, Chanhyuk Lee, Donggyun Kim et al.

FlowSeek integrates depth foundation models and motion bases for low-resource, high-accuracy optical flow estimation, outperforming SOTA by 10-15%.

Matteo Poggi, Fabio Tosi

Proposes MDT-dist, a few-step flow distillation method reducing sampling from 25 to 1-2 steps, enabling real-time 3D generation.

Zanwei Zhou, Taoran Yi, Jiemin Fang et al.

OneCAT is a decoder-only autoregressive model for unified multimodal understanding, generation, and editing, eliminating external encoders for efficiency.

Han Li, Xinyu Peng, Yaoming Wang et al.

Kwai Keye-VL 1.5 employs Slow-Fast video encoding, extends context to 128K tokens, combined with multi-stage pretraining and post-training, greatly enhancing video understanding.

Biao Yang, Bin Wen, Boyang Ding et al.

FantasyHSI enables 4D human synthesis in any scene using a graph-based multi-agent framework, significantly improving task completion.

Lingzhou Mu, Qiang Wang, Fan Jiang et al.

POINTS-Reader employs a two-stage, distillation-free framework using synthetic data and self-improvement, achieving state-of-the-art document conversion performance.

Yuan Liu, Zhongyin Zhao, Le Tian et al.

GPSToken employs Gaussian parameterization for spatially adaptive image tokenization, achieving state-of-the-art FID=1.50 in image generation with 128 tokens.

Zhengqiang Zhang, Rongyuan Wu, Lingchen Sun et al.

VoxHammer is a training-free 3D editing method using inverse trajectory prediction and feature replacement.

Lin Li, Zehuan Huang, Haoran Feng et al.

Drawing2CAD employs sequence-to-sequence learning to convert SVG vector drawings into parametric CAD operation sequences.

Feiwei Qin, Shichao Lu, Junhao Hou et al.

JCo-MVTON leverages multi-modal diffusion transformers for mask-free virtual try-on, achieving SOTA performance on DressCode dataset.

Aowen Wang, Wei Li, Hao Luo et al.

UniGen framework uses CoMoE module and WeaveNet mechanism for efficient controllable image generation, excelling on Subjects-200K and MultiGen-20M datasets.

Guoqing Zhang, Xingtong Ge, Lu Shi et al.

MV-RAG integrates retrieval with multiview diffusion, significantly improving out-of-domain concept 3D generation with 15%+ performance boost.

Yosef Dayani, Omer Benishu, Sagie Benaim