The DragGaussian framework leverages 3D Gaussian Splatting for interactive point-based 3D object editing. Users upload pre-trained 3D Gaussian object models through a user interface, enabling them to perform drag-based editing by selecting starting and ending points or editable regions. These points are projected onto 2D images from different camera perspectives, facilitating multi-view consistent editing. To enhance adaptability, the pre-trained 2D editing network undergoes fine-tuning using an improved version of multi-view LoRA. The refined 2D images are then used to modify the original 3D Gaussian model, showcasing the edited appearance of the 3D objects.

The DragGaussian pipeline encompasses various stages, including interactive 3D point-based manipulation, multi-view identity preserving fine-tuning, and multi-view consistent editing. The interactive manipulation allows users to select anchor and destination points for editing, while the fine-tuning process enhances the adaptability of the pre-trained 2D editing network. Multi-view consistent editing involves motion supervision, point tracking, and multi-view consistent denoising to generate modified 2D images that accurately reflect the desired 3D object edits. The framework's user interface facilitates intuitive drag-based editing, enabling precise and efficient modifications to 3D Gaussian object models.

By combining the efficiency of 3D Gaussian Splatting with diffusion models, DragGaussian introduces a novel approach to 3D object editing. The framework's emphasis on interactive point-based manipulation and multi-view consistent editing sets it apart from existing methods, offering users a user-friendly and effective tool for precise 3D editing tasks. Through comprehensive qualitative and quantitative experiments, the effectiveness of DragGaussian in enabling drag-style manipulation on 3D Gaussian representations is demonstrated, highlighting its potential for enhancing the user experience in 3D object editing.