Research Papers

WorldGrow redefines 3D world generation by enabling infinite, continuous 3D scene creation through a hierarchical block-wise synthesis and inpainting pipeline. Developed by researchers from Shanghai Jiao Tong University, Huawei Inc.,

Nano3D revolutionizes 3D asset editing by enabling training-free, part-level shape modifications like removal, addition, and replacement without any manual masks. Developed by researchers from Tsinghua University, Peking University, HKUST, CASIA,

Triangle Splatting+ redefines 3D scene reconstruction and rendering by directly optimizing opaque triangles, the fundamental primitive of computer graphic,  in a fully differentiable framework. Unlike Gaussian Splatting or NeRF-based approaches,

Code2Video introduces a revolutionary framework for generating professional educational videos directly from executable Python code. Unlike pixel-based diffusion or text-to-video models, Code2Video treats code as the core generative medium, enabling

CAP4D introduces a unified framework for generating photorealistic and animate style rendering 4D portrait avatars from any number of reference images as well as even a single image. By combining

Test3R is a novel and simple test-time learning technique that significantly improves 3D reconstruction quality. Unlike traditional pairwise methods such as DUSt3R, which often suffer from geometric inconsistencies and poor generalization,

BlenderFusion is a novel framework that merges 3D graphics editing with diffusion models to enable precise, 3D-aware visual compositing. Unlike prior approaches that struggle with multi-object and camera disentanglement, BlenderFusion

LongSplat is a new framework that achieves high-quality novel view synthesis from casually captured long videos, without requiring camera poses. It overcomes challenges like irregular motion, pose drift, and memory

DINOv3 is a next-generation vision foundation model trained purely with self-supervised learning. It introduces innovations that allow robust dense feature learning at scale with models reaching 7B parameters and achieves

Genie 3 is a general-purpose world model which, given just a text prompt, generates dynamic, interactive environments in real time and rendered at 720p, 24 fps, while maintaining consistency over

PartCrafter is the first unified 3D generative model that jointly synthesizes multiple semantically meaningful and geometrically distinct 3D parts from a single RGB image without any segmentation required. Powered by

SimLingo unifies autonomous driving, vision-language understanding, and action reasoning-all from camera input only. It introduces Action Dreaming to test how well models follow instructions, and outperforms all prior methods on