A possible bug in pipeline_ddpm.py with start_t

[ccj5351]

Hello,

Thanks for sharing the code. There is a possible bug for your check:

if start_t is None:
            model_input = torch.cat([inputs[:, :6], image], dim=1)
            timesteps = self.scheduler.timesteps
        else:
            model_input = inputs
            timesteps = self.scheduler.timesteps[-start_t:]
            # need your check
            # since we do not want to use a random noise for start_t being
            # a non-zero step (for example: the coarse-to-fine pattern)
            image = inputs[:,-2:,:,:] # use coarse flow you predicted already

[ccj5351]

This change will help improve the accuracy when running evaluate_diffusers_tile.py.