• A Diffusion model for indoor scene synthesis

Notes

• 3D Scene $s$ → fully-connected scene graph $x_0$
  • each object as a graph node strong all object attributes, i.e. location, size, orientation, class label, and latent shape code.
• DiffuScene → Based on a sel of all posible $x_0$
  • Forward: Gradually add noice to $x_0$ until a standard Gaussian noice $x_T$
  • Reverse: A Denoising network cleans the noisy graph using ancestral sampling
    • use the denoised object features to perform shape retrieval

Graph Representation

Denoising Network

Text-Conditioned

1. employ a pretrained BERT encoder to extract word embedings
2. inject the language guidance into denoising network using cross attention layers