the output of controlnet tile is awasome, and i want to train a tile for custom dataset. However, the resultis strange.
my training setup:
use the train_controlnet.py script from diffusers.
for the high-resolution image with shape (width, height), we downsample hr img to (width//k, height//k), where k is random sample from {2, 4, 8}, then upsample to (width, height) as controlnet input. and hr img as sd input. I train cn with lr 1e-5 for 20k update steps.
however, the result seems bad, especially in the x4 x8 case, the output appears as a grid. Can you point out where my training Settings are wrong?
the output of controlnet tile is awasome, and i want to train a tile for custom dataset. However, the resultis strange. my training setup: use the train_controlnet.py script from diffusers.
for the high-resolution image with shape (width, height), we downsample hr img to (width//k, height//k), where k is random sample from {2, 4, 8}, then upsample to (width, height) as controlnet input. and hr img as sd input. I train cn with lr 1e-5 for 20k update steps.
however, the result seems bad, especially in the x4 x8 case, the output appears as a grid. Can you point out where my training Settings are wrong?
x2 upsample![step200000_dog thumb2](https://github.com/lllyasviel/ControlNet-v1-1-nightly/assets/28671582/80b91b8a-255b-4092-95ea-40fb380657d2)
x4 upsample![step200000_dog thumb4](https://github.com/lllyasviel/ControlNet-v1-1-nightly/assets/28671582/8b50f0c1-9658-43ba-8edf-2994309ca118)
x8 upsample![step200000_dog thumb8](https://github.com/lllyasviel/ControlNet-v1-1-nightly/assets/28671582/3f7edd60-43eb-4f6b-9c93-ab0b04a0f782)