What about the performance on the MNIST dataset

[1999kevin]

What about the performance on the MNIST dataset？ And waht about the GPU comsumption?

[thorinf]

I'm updating the readme now. Here is an output after 100 epochs:

If you add use the same noise for every iteration at sampling then the images do not switch between different numbers as much. This is the intended property of 'self-consistency' from the paper. Generations on the same trajectory producing the same sample. However the multi-step sampler algorithm in the paper does add different noise each iteration. Really up to you which you prefer.

The GPU consumption is pretty low, I have only 2GB of VRAM. It could probably be a lot better with a smaller NN - I think the UNet I've used is too big.

[1999kevin]

Thanks for your quick answers. The performance is pretty good. I find that you use L2 loss. I guess this will also reduce the GPU consumption

[thorinf]

Here is sampling where the same noise is injected at every iteration.

Yeah I think L2 is fine for this task. The goal is the implementation is A) for me to learn, and B) create something minimal.

[1999kevin]

So appreciate of your implementation on MNIST!

[thorinf]

Please give the project a watch or a star if you can. Currently looking for work and a bit of Github exposure might help.

[1999kevin]

Sure, I'm glad to help.