Open chenerg opened 4 weeks ago
The generative UV features plane is learnable parameters with a shape of (32, 256, 256) for each subject. It is initialized to a random code and then optimized during fitting and denoising. Therefore, you can add Gaussian noise to the feature plane during denoising. If you still have questions, please feel free to ask.
Thanks for your work. But I have a question while reading the training part of the paper. It said
Specifically, we first add Gaussian noise 𝜖 ∼ N (0, 𝐼) into the given generative UV features plane 𝑥𝑖 via a noise schedule comprising differentiable functions 𝛼 (𝑡) and 𝜎(𝑡), obtaining a noisy feature plane 𝑥𝑖(𝑡) := 𝛼 (𝑡)𝑥𝑖 + 𝜎(𝑡)𝜖 at diffusion time step 𝑡. T
, however, at first we don't have any given UV features (The paper said they should be trained with the denoising training process), so what should the Gaussian noise be added to?