This PR aims at generalizing the three pipeline examples to make them work with the schedulers that need to scale the denoising model input depending on the current timestep and that have a initial noise standard deviation that depends on the timesteps.
This is done making all of them implementing the following methods
init_noise_sigma
scale_model_input
The pipeline examples are modified as follows:
the initial latents are rescaled by the scheduler's init_noise_sigma
the latents are first passed to the scheduler's scale_model_input method before being fed to the unet forward pass.
This PR aims at generalizing the three pipeline examples to make them work with the schedulers that need to scale the denoising model input depending on the current timestep and that have a initial noise standard deviation that depends on the timesteps.
This is done making all of them implementing the following methods
init_noise_sigma
scale_model_input
The pipeline examples are modified as follows:
init_noise_sigma
latents
are first passed to the scheduler'sscale_model_input
method before being fed to the unetforward
pass.