How to modify the code to adapt to smaller size of input?

[abcdvzz]

Hi there,

Thank you for your amazing work. The thing is I don't have A6000 GPUs at hand. So I modified the input size to 320x320 to adapt to the GPU memory. However, the result is terrible. I would like to know if there are specific settings I need to modify in the code. It seems that it should not relate to the input size. I have modified the 'int(512y1):int(512y2), int(512x1):int(512x2)' in the plms.py. Do I need to modify the 'v1-inferency.ymal' about stable diffusion? I also modify the learning rate in plms.py. Nothing works. Could you please tell me if there is anything I need to pay attention to?

Thank you very much.

[wuqiuche]

Hi abcdvzz,

Thank you for your interests in our work and sorry for the late reply. We have updated the code, which should have better efficiency, smaller memory cost (but still ~40 GB), and clearer structure. I believe "plms.py" should be the only script that needs to be changed. However, I guess it would also be a good idea to check the mask positions when changing the resolution. For this purpose, we implement a "plot" function within "ldm/modules/attention.py", which allows you to check the intermediate mask position and make sure they are at the correct positions after reshaping.