SPHINX-X scales data and parameters, achieving state-of-the-art multi-modal performance across diverse tasks, from 1.1B to 56B parameters.

Dongyang Liu, Renrui Zhang, Longtian Qiu et al.

UAV-Rain1k dataset uses Blender-based rain modeling, evaluates six deraining models, revealing Transformer-based methods outperform CNNs in high-res UAV imagery.

Wenhui Chang, Hongming Chen, Xin He et al.

Proposes time-varying correlated blue noise in diffusion models, improving image quality with 5-10% lower FID than Gaussian noise.

Xingchang Huang, Corentin Salaün, Cristina Vasconcelos et al.

InstructScene integrates semantic graph prior with diffusion and transformer decoders, achieving 73.64% iRecall and superior scene control on 3D-FRONT.

Chenguo Lin, Yadong Mu

ConsistI2V enhances image-to-video generation consistency using diffusion models, showing significant improvements in experiments.

Weiming Ren, Huan Yang, Ge Zhang et al.

CogCoM uses Chain-of-Manipulations reasoning for precise visual reasoning, achieving state-of-the-art across 9 benchmarks with 17B parameters.

Ji Qi, Ming Ding, Weihan Wang et al.

MobileVLM V2 employs novel architecture and high-quality data, with 1.7B parameters, outperforming comparable 3B models in accuracy and speed.

Xiangxiang Chu, Limeng Qiao, Xinyu Zhang et al.

Decoupled camera and object motion control in text-to-video via spatial and temporal cross-attention, without extra optimization.

Shiyuan Yang, Liang Hou, Haibin Huang et al.

DeCoF leverages frame consistency to detect AI-generated videos, achieving 92% accuracy on unseen models and demonstrating strong generalization.

Long Ma, Zhiyuan Yan, Qinglang Guo et al.

UniTouch aligns tactile embeddings with pretrained image models, enabling zero-shot multimodal tactile tasks with high accuracy.

Fengyu Yang, Chao Feng, Ziyang Chen et al.

PLoRA adapts only image tokens, enabling InternLM2-7B to compose and understand free-form interleaved text-image content.

Xiaoyi Dong, Pan Zhang, Yuhang Zang et al.

Motion-I2V combines diffusion-based motion prediction with an enhanced temporal attention mechanism to produce consistent, controllable videos from images, supporting sparse trajectories and zero-shot translation.

Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang et al.

Study shows Soft MoE outperforms sparse MoE under fixed compute budget.

Tianlin Liu, Mathieu Blondel, Carlos Riquelme et al.

Grounded SAM combines Grounding DINO with SAM, achieving 48.7 mAP on SegInW zero-shot benchmark, enabling open-vocabulary detection and segmentation.

Tianhe Ren, Shilong Liu, Ailing Zeng et al.

Proposed Energy-Based Concept Bottleneck Models, improving prediction accuracy and concept interpretation.

Xinyue Xu, Yi Qin, Lu Mi et al.

LAA-Net: Localized Artifact Attention Network achieves 99.96% AUC in high-quality deepfake detection by explicitly focusing on local artifacts using multi-task learning and enhanced feature pyramid networks.

Dat Nguyen, Nesryne Mejri, Inder Pal Singh et al.

Proposed HAZARD benchmark evaluates embodied agents' decision-making in dynamic scenarios with LLMs, achieving up to 77.8% rescue value in fire, flood, and wind environments.

Qinhong Zhou, Sunli Chen, Yisong Wang et al.

SpatialVLM employs large-scale internet data to enhance vision-language models' 3D spatial reasoning, achieving 99% accuracy in distance estimation.

Boyuan Chen, Zhuo Xu, Sean Kirmani et al.

Introduced Missing Modality Token (MMT) to reduce performance drop to 10% on Ego4D dataset.

Merey Ramazanova, Alejandro Pardo, Humam Alwassel et al.

Vlogger combines LLM and diffusion models to generate 5-minute coherent long videos, maintaining script-actor consistency.

Shaobin Zhuang, Kunchang Li, Xinyuan Chen et al.