SAM2Matting leverages pre-trained VOS trackers (e.g., SAM2, SAM3) with region proposal and multi-scale alpha refinement, achieving state-of-the-art zero-shot video and image matting performance.

Ruiqi Shen, Guangquan Jie, Chang Liu et al.

SimPhysNet combines physics-informed neural networks with self-supervised contrastive learning to achieve 96.06% accuracy in laser welding penetration prediction using only 200 labeled images.

Sen Li, Xiaoying Liu, Xiaojian Xu et al.

FLUX3D employs Diffusion-Aligned Structured Latents and sparse-structure-aware diffusion transformer to generate high-fidelity 3D Gaussian point clouds, outperforming SOTA methods.

Haorui Ji, Weizhe Liu, Hongdong Li et al.

Proposes a semantic browsing framework for controllable diversity in image generation, leveraging rich textual representations to enable structured, interpretable exploration.

Sara Dorfman, Maya Vishnevsky, Omer Dahary et al.

JanusMesh is a fast, zero-shot framework for generating dual-semantic 3D illusions using cross-space denoising, completing in 3-5 minutes with high geometric and semantic fidelity.

Siang-Ling Zhang, Huai-Hsun Cheng, Tsung-Ju Yang et al.

UNIEGO employs proxy-mediated hierarchical distillation from nine heterogeneous teachers to unify egocentric video representations, achieving state-of-the-art results.

Wenhao Chi, Arkaprava Sinha, Dominick Reilly et al.

Proposes Spatially Speculative Decoding (SSD), leveraging 2D spatial prediction to accelerate autoregressive image generation by up to 13.3×.

Shilong Xiang, Zirui Zhang, Lijun Yu et al.

CalTennis is a large multi-view tennis video dataset with over 11 million frames, used to evaluate monocular-to-3D pose estimation, highlighting challenges in depth and foot contact accuracy.

Ilona Demler, Xinran Xie, Blake Werner et al.

UniAR introduces a unified autoregressive framework with a single discrete visual tokenizer, achieving state-of-the-art results in image generation and understanding.

Wujian Peng, Lingchen Meng, Yuxuan Cai et al.

EventDrive integrates event cameras with vision-language models, significantly improving perception, understanding, prediction, and planning in autonomous driving.

Dongyue Lu, Rong Li, Ao Liang et al.

Proposes Qwen-RobotWorld, a language-conditioned video world model using double-stream MMDiT and 8.6M embodied video-text pairs, achieving top performance on multiple benchmarks.

Jie Zhang, Xiaoyue Chen, Anzhe Chen et al.

MeshLoom is a feed-forward non-rigid mesh registration network that reconstructs vertex deformations across sequences within seconds, outperforming state-of-the-art methods.

Jianqi Chen, Jiraphon Yenphraphai, Xiangjun Tang et al.

This study identifies a small set of attention heads—gaze heads—in VLMs that causally track the current description region, enabling effective inference-time control via attention masks.

Rohit Gandikota, David Bau

Introduces OmniVideo-100K, a large-scale dataset with structured scripts and evidence chains, boosting audio-visual reasoning by up to 20.59%.

Xinyue Cai, Chaoyou Fu, Yi-Fan Zhang et al.

Introducing RATS (Register Attention Transformers), which self-supervisedly discovers part-level structures with N learnable registers, achieving +12 mIoU on five segmentation benchmarks.

Timing Yang, Predrag Neskovic, Jansen Seheult et al.

Instruct-Particulate employs large-scale heterogeneous datasets and instruction-guided neural networks to efficiently predict 3D articulated structures, significantly improving generalization.

Ruining Li, Yuxin Yao, Matt Zhou et al.

InterleaveThinker employs a multi-agent framework with a planner and critic, achieving high-quality interleaved text-image generation with step-wise reinforcement learning, improving performance on benchmarks by over 50%.

Dian Zheng, Harry Lee, Manyuan Zhang et al.

Proposes Modality Forcing, a post-training method enabling a single DiT model to jointly generate image and sparse depth data, achieving 57% reduction in AbsRel and scaling with model size.

Bardienus Pieter Duisterhof, Deva Ramanan, Jeffrey Ichnowski et al.

SpatialClaw employs code as an action interface, achieving 59.9% average accuracy across 20 spatial reasoning benchmarks, outperforming recent models by 11.2%.

Seokju Cho, Ryo Hachiuma, Abhishek Badki et al.

Flex4DHuman employs relative camera-pose encoding within a diffusion framework to synthesize synchronized multi-view videos from monocular or sparse inputs, surpassing prior methods without explicit geometry priors.

Jen-Hao Cheng, Yipeng Wang, Hao Zhang et al.