Demo, which originally required 22GB of VRAM, now got leaner and can actually run in Colab on a single Tesla T4. This means you can explore large-scale diffusion models for novel view generation or 3D object reconstruction on less expensive hardware.

For an overall reduced memory footprint, inference is done with Accelerate, using sharded checkpoints of the underlying large-scale diffusion model.

Colab notebook

Sharded model checkpoints