Open weleen opened 4 months ago
Hi, thanks for your issue!
The commands look right to me — how long are you training for (and on what hardware)?
Those commands should produce a model that is slightly (but not much much) worse than an equivalent DiT model when sampling for 250 steps (because the DiT model has so many more parameters, and you’re sampling for so many steps). The results you posted look strange — the balloon class for example should give balloons that look ok-ish even toward the beginning of training.
For the pretrained models, I’m sorry! I meant to release them but got caught up in other things. I’ll get on that.
Hope this helps and I’ll try to update with the pretrained models soon!
Best, Luke
Hi, thanks for your issue!
The commands look right to me — how long are you training for (and on what hardware)?
Those commands should produce a model that is slightly (but not much much) worse than an equivalent DiT model when sampling for 250 steps (because the DiT model has so many more parameters, and you’re sampling for so many steps). The results you posted look strange — the balloon class for example should give balloons that look ok-ish even toward the beginning of training.
For the pretrained models, I’m sorry! I meant to release them but got caught up in other things. I’ll get on that.
Hope this helps and I’ll try to update with the pretrained models soon!
Best, Luke
Dear Luke,
Thank you for your prompt response.
I trained the model on 8 GTX 3090 for 600k steps, I also run sample.py by adopting the model trained via fast-DiT, it looks good.
The generated images are really wired, do you have any idea about this issue?
Best Regards, Yiming
Hi,
Thanks for your open-sourced code, I try to reproduce the results of fpdm on ImageNet, but I find the generated images are much worse than original fast-DiT model. Could you help me check if I run the code correctly?
I try to train model with fixed-point layers and disable zero_snr/v_pred, the arguments are as follow:
then test by running
the testing arguments are as follows:
I take some samples:
![00006--417--balloon](https://github.com/lukemelas/fixed-point-diffusion-models/assets/7327734/a6399451-4a13-4359-9b0a-a3b75e163b4a)
Additionally, when do we expect to have the pre-trained model?
Best Regards