shap-e
shap-e copied to clipboard
why applying Gaussian diffusion noise when converting latents to MLP weights
Hi, in your paper, I find that "after computing the sequence of latents, we apply Gaussian diffusion noise q(ht) to the latents with very little noise. After the noise and bottleneck layers, we project each latent vector to 256 dimensions probability 0.1. For this diffusion noise, we use the schedule $\hat{\alpha} = 1 − t^5$ which typically produces and stack the resulting latents into four MLP weight matrices of size 256 × 256.
I wonder why is this necessary? Does it help to improve the robustnesss of INR?