Question about TCD(1)

[zjhJOJO]

Thank you for sharing your code. I have a question regarding a particular aspect that I found a bit confusing.
In the DPM-solver paper, the authors demonstrate that DPM-solver(1) is identical to DDIM. However, in your manuscript, you mainly employ DPM-solver(1) to reparameterize the consistency function in CM and LCM (which is parameterized by DDIM), leading to an improvement in few-step generation. I am having trouble understanding how this is achieved.
Could you kindly provide some clarification on this point? I would greatly appreciate your insight.

[mhh0318]

in our paper we reparameter the prediction to a smaller step instead of origin, which also enable a better TCD sampler to avoid large stochastic and discretization.