We present SAM 3D, a generative model for visually grounded 3D object reconstruction, predicting geometry, texture, and layout from a single image.
Personal Contribution: Main contributor to post-training, preference optimization.

After training on large video data, will global 3D understanding naturally emerge? We study this by quantifying the 3D understanding of existing Video Foundation Models pretrained on vast video data, by estimating multiple 3D properties from their features via shallow read-outs.

We propose an agentic framework for text-guided 3D editing. Vinedresser3D uses a multimodal LLM to interpret editing prompts and performs precise, mask-free 3D editing directly in the latent space of a native 3D generative model.

Which image cues do 3D generation models actually rely on? We systematically perturb monocular cues (shading, texture, silhouette, perspective, etc.) across seven state-of-the-art methods and quantify their impact on 3D generation.

Symmetry is a ubiquitous property of real-world objects, yet 3D generation models ignore this. We introduce Reflect3D, a zero-shot 3D reflection symmetry detector from a single image, and show that incorporating detected symmetry into 3D generation improves structural accuracy and visual fidelity.

Video object segmentation (VOS) models struggle when objects change state (e.g., an egg being cracked). We introduce the VSCOS benchmark for this underexplored setting, and propose an improved baseline for VOS of objects undergoing state changes.

We introduce YouTubePD, the first public multimodal benchmark for Parkinson's Disease (PD) analysis, crowdsourced from existing YouTube videos featuring over 200 subjects.

In few-shot segmentation, class diversity matters more than data quantity. We introduce FSS-1000, a 1000-class dataset with pixel-wise annotations that prioritizes breadth over depth, enabling more robust generalization and more precise evaluation.