This paper introduces layer magnitude preservation and post-hoc EMA tuning, reducing ImageNet-512 FID from 2.41 to 1.81, significantly improving diffusion model training stability and quality.

Tero Karras, Miika Aittala, Jaakko Lehtinen et al.

This study questions whether ego status alone suffices for open-loop end-to-end autonomous driving, revealing dataset bias and metric limitations that favor ego-only reliance.

Zhiqi Li, Zhiding Yu, Shiyi Lan et al.

TimeChat is a time-sensitive multimodal large language model improving long video understanding, notably achieving a 2.8 CIDEr increase on YouCook2.

Shuhuai Ren, Linli Yao, Shicheng Li et al.

Proposes SCLIP with Correlative Self-Attention, boosting zero-shot semantic segmentation from 14.1% to 38.2% mIoU.

Feng Wang, Jieru Mei, Alan Yuille

VideoBooth enables feed-forward image-prompted video generation, reaching 65.0979 DINO and 74.7971 CLIP-Image.

Yuming Jiang, Tianxing Wu, Shuai Yang et al.

DeepCache accelerates diffusion models by caching features, achieving 2.3× speedup on Stable Diffusion v1.5 with minimal quality loss without retraining.

Xinyin Ma, Gongfan Fang, Xinchao Wang

DREAM integrates diffusion rectification and estimation adaptation, boosting training speed (2-3×) and reducing sampling steps (10-20×) in super-resolution tasks.

Jinxin Zhou, Tianyu Ding, Tianyi Chen et al.

GraphDreamer uses scene graphs to generate disentangled 3D scenes, outperforming prior methods with CLIP score 0.3326.

Gege Gao, Weiyang Liu, Anpei Chen et al.

PI-ViT integrates 2D/3D skeleton info during training to enhance video transformer-based activity recognition, achieving SOTA without extra inference cost.

Dominick Reilly, Srijan Das

ChatPose employs multimodal LLMs with embedded SMPL parameters for 3D human pose understanding and generation.

Yao Feng, Jing Lin, Sai Kumar Dwivedi et al.

Distribution Matching Distillation (DMD) converts diffusion models into one-step generators, achieving FID 2.62 on ImageNet 64×64 at 20 FPS.

Tianwei Yin, Michaël Gharbi, Richard Zhang et al.

X-InstructBLIP aligns multi-modal data to LLMs, enabling cross-modal reasoning with significant performance gains.

Artemis Panagopoulou, Le Xue, Ning Yu et al.

DiffCAD employs diffusion models for weakly-supervised CAD retrieval and alignment from RGB images, surpassing supervised SOTA with synthetic training.

Daoyi Gao, Dávid Rozenberszki, Stefan Leutenegger et al.

MGProto uses Gaussian-distributed prototypes for interpretable and trustworthy image recognition, excelling on CUB-200-2011.

Chong Wang, Yuanhong Chen, Fengbei Liu et al.

HOLD is a category-agnostic monocular method for joint 3D reconstruction of hands and objects, outperforming supervised baselines.

Zicong Fan, Maria Parelli, Maria Eleni Kadoglou et al.

Ego-Exo4D leverages multimodal multiview video datasets for fine-grained human activity understanding, achieving 85% accuracy in activity recognition.

Kristen Grauman, Andrew Westbury, Lorenzo Torresani et al.

Drive-WM introduces a driving world model for controllable multiview future generation and planning.

Yuqi Wang, Jiawei He, Lue Fan et al.

VITATECS dataset evaluates video-language models' temporal concept understanding, revealing deficiencies in temporal comprehension.

Shicheng Li, Lei Li, Shuhuai Ren et al.

LightGaussian achieves 15× compression and boosts FPS to 237 by pruning, distilling, and quantizing 3D Gaussians.

Zhiwen Fan, Kevin Wang, Kairun Wen et al.

UniIR employs multi-task training and instruction tuning, achieving robust multi-modal retrieval across 8 tasks with 10 datasets, showing strong zero-shot generalization.

Cong Wei, Yang Chen, Haonan Chen et al.