Cheng Lu

Here are some of my own suggestions: 1. For few-step guided sampling, I still recommend using multi-step DPM-Solver++. 2. For obtaining a converged sample with a faster convergence rate, I...

Here are some examples for comparing multi-step and single-step DPM-Solver++ (2nd order, midpoint):    

No problem. I will add tests and flax version implementations :)

Hi @qsh-zh , The comparison seems not fair because I used your `t-AB-DEIS` in my paper :( Is the `log-rho-DEIS` better than `t-AB-DEIS` or DPM-Solver? (On CIFAR-10 it seems that...

Hi @qsh-zh , Thank you for modifying the message. I'm sorry that I find the above comments I left seem to be too rude. I must say I apologize for...

Hi guys, Hope you are doing well. Happy to know that you all approve my implementation for the multi-step DPM-Solver & DPM-Solver++. I'm also thinking about the implementation of single-step...

Hi guys @keturn @hafriedlander @patrickvonplaten @pcuenca , I've implemented a pytorch version of single-step DPM-Solver in this PR: https://github.com/huggingface/diffusers/pull/1442 Is it good for you? Looking forward to further discussions!

In addition, I even think that we may unify the implementations of all the schedulers because it seems that they all can be understood as the above process, and the...

I think one of the reasons is that `k-diffusion` uses a clean and easy wrapper for the whole solver. Maybe we can also wrap the schedulers into **one** file by...

Please check this example code:https://github.com/LuChengTHU/dpm-solver/tree/main/examples/ddpm_and_guided-diffusion