The clip_sample option for the DDPM and DDIM schedulers clip the predicted noise between -1 and 1. This procedure was borrowed from the SD implementation, and it should be allowed for the user to specify them; for latent diffusion models, where ranges are ~Gaussian, clipping can hinder the training process. This is also in response to Discussion #492: https://github.com/Project-MONAI/GenerativeModels/discussions/492.
The clip_sample option for the DDPM and DDIM schedulers clip the predicted noise between -1 and 1. This procedure was borrowed from the SD implementation, and it should be allowed for the user to specify them; for latent diffusion models, where ranges are ~Gaussian, clipping can hinder the training process. This is also in response to Discussion #492: https://github.com/Project-MONAI/GenerativeModels/discussions/492.