improved-diffusion
improved-diffusion copied to clipboard
Mistake in accelerated DDIM sampling?
When I'm debugging the image sampler with fewer timesteps (let's say just 2 steps out of 1000), in .ddim_sample(), in the first sampling step, the variable t is set to all ones, which the _WrappedModel translates to 500, and calls the actual model with x_T (which is totally random) but time step conditioning 500 (in the variable new_ts).
In my understanding, the model should be called with timestep 1000 instead. Am I understanding something wrong or is there a mistake in the implementation?