Problem while trying to generate 1024 by 1024 image

[dfav22]

Hi, when I try to use this method (on Forge webui) with 1024x1024 resolution (needed by SDXL) it throws this error:

File ".../extensions/CharacteristicGuidanceWebUI/scripts/CharaIte.py", line 361, in compute_correction_direction
dxs_cond_part = torch.cat( [*( [(h - 1) * dxs[:,None,...]]*num_x_in_cond )], axis=1 ).view( (dxs.shape[0]*num_x_in_cond, *dxs.shape[1:]) )
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: torch.cat(): expected a non-empty list of Tensors
torch.cat(): expected a non-empty list of Tensors

Which is caused by num_x_in_cond being zero. Any ideas on how to fix that?

[scraed]

Hi, I can't reproduce this issue on 1024*1024 image. Could you provide the batch count, batch number, and the prompt (especially special tokens) you are using?

[dfav22]

Hi,I use batch number of two and batch size of one. My prompt is: "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k".

[dfav22]

Maybe it has to do something with low vram flag I use?

[scraed]

yes, I think it is related the low vram flag. Seems the positive and negative peompts aren't envalued simultaneously in this case. But I still can't reproduce it even with low vram flag on my machine. May I ask how large vram do you have? I might need to impose some hard vram limit on my machine to reproduce it.

PS. It would nice if you could print out the shape of the tensor x_in, dxs and uncond during evaluation. As num_x_in_cond is computed from their shape by
num_x_in_cond = len(x_in[:-uncond.shape[0]])//len(dxs)

[dfav22]

Hi, my VRAM size is 4 GB.

Here are sizes which cause an error:
UNCOND:torch.Size([1, 77, 2048])
X_IN:torch.Size([1, 4, 128, 128])
DXS:torch.Size([1, 4, 128, 128])

Notably, running this on CPU makes it work. Here are sizes which are printed out while running on cpu:
UNCOND:torch.Size([1, 77, 2048])
X_IN:torch.Size([2, 4, 128, 128])
DXS:torch.Size([1, 4, 128, 128])

[scraed]

Bingo! X_IN:torch.Size([1, 4, 128, 128]) is causing the problem. The first dimension is usually 2. I think it is the vram saving mechanism of forge. Will try to figure out a way to fix it.