Computer Vision

New method for moving object segmentation by combining long-range motion cues, semantic features, and SAM2, achieving state-of-the-art performance in challenging scenarios.

ORIGEN: First zero-shot 3D orientation grounding in text-to-image generation.

Hi3DGen: High-fidelity 3D geometry generation from images via normal bridging.

X2-Gaussian enables continuous-time 4D CT reconstruction via dynamic radiative Gaussian splatting and self-supervised respiratory motion learning.

VBench 2.0: A new benchmark suite advancing video generation evaluation with intrinsic faithfulness metrics.

SparseFlex: Achieves high-res, arbitrary-topology 3D shape modeling via sparse isosurface representation and sectional voxel training. Revolutionizing 3D generative AI!

HSMR: Reconstructing 3D humans with a biomechanically accurate skeleton model from a single image, enhancing pose realism.

Optimal Stepsize Distillation accelerates diffusion sampling by distilling knowledge from reference trajectories, achieving 10x speedup with minimal performance loss.

Lumina-Image 2.0: A unified & efficient image generative framework, outperforming previous models with only 2.6B parameters.

LOCATEdit refines cross-attention maps with graph Laplacian regularization, achieving precise & localized text-guided image editing without artifacts.

LeX-Art: High-quality text-to-image generation via scalable data synthesis.

This survey explores the evolution of physics cognition in video generation, addressing the gap between visual realism and physical accuracy.

ChatAnyone: Stylized real-time portrait video generation with hierarchical motion diffusion model.

Fixing fine-tuned diffusion models! By using richer, unconditional priors, they generate better images and videos.

Synthetic data can enhance the physical realism of video synthesis, paving the way for more believable generated content.

Free4D: Tuning-free 4D scene generation with spatial-temporal consistency.

Feature4X: 4D Agentic AI from Monocular Video w/ Gaussian Feature Fields

DINeMo: Learns 3D models with no 3D annotations, leveraging pseudo-correspondence from visual foundation models for enhanced pose estimation.

BIZGEN: Article-level Visual Text Rendering for Infographics Generation

LongTextAR advances long-text image generation via a novel tokenizer, enabling accurate, controllable, and high-fidelity text rendering in images.