Fig. 4
From: Generation driven understanding of localized 3D scenes with 3D diffusion model
The 3D-UDDPM. The Markov chain is complete, beginning with a local cube devoid of noise and subsequently injecting inhomogeneous Gaussian noise while embedding the time step \(\text {t}\) until the entire chain is completely injected with noise. Reverse inference is then performed to estimate the noise \(\varepsilon _{\theta }(x_{t},t)\) at a time step using the learned noise distribution, gradually recovering normal, clearly localized three-dimensional objects.
