The pipeline removes the background, conditions generation on image features, and uses a diffusion transformer and decoder to produce Gaussian representations. Users can vary sampling settings and request between 32,768 and 1,048,576 Gaussians, then save the result as a standard PLY splat.
IDOL is useful for quickly exploring 3D object concepts from existing visual references. Its Python interface and Gradio demo support experimentation with asset density, input type, and reproducible seeds before bringing generated splats into a compatible viewer or downstream workflow.
