Have you reproduced similar results as original paper?

[lyc0929]

I have been training with a single A100 for four days and have achieved significant results, which makes me convinced that the code works.In addition, I sent you an email, please check it.

[siqi0905]

I'm also curious about how you reproduced the same results as the original paper. Could you share more training details? Also, I'm wondering if training only the main UNet is feasible.

[xuyu666]

Hello, author I'd also like to see the results of your reproduction, can you send me a copy of my e-mail, my e-mail is 2622960263@qq.com

[ManuelCecere]

Hello author, would you be available to share your results with me too? My email is manucp44@gmail.com I was wondering about the loss function you wrote: loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")

since it seems that noise is defined as torch.randn_like(latents) so I don't get how the model is learning, could you help me understand that?

[appleyang123]

I use your code. But after training 40k steps, I have tested model on test data. The results are here:

[orshi]

I have been training with a single A100 for four days and have achieved significant results, which makes me convinced that the code works.In addition, I sent you an email, please check it. also expected to see the results of your reproduction it_xie@hotmail.com