Can't reproduce the result in fill50k

[JunnYu]

Hi, i can't reproduce the result in fill50k, https://github.com/lllyasviel/ControlNet/blob/main/docs/train.md. i use the same Hyperparameter in the https://github.com/lllyasviel/ControlNet/blob/main/tutorial_train.py Could you help me to reproduce the result? Thx, here is my generated images after 4000 training steps.

[lllyasviel]

is that sd1 or sd2? sd1 will fit condition at about 4k to 6k iter. sd2 is usually at 6k to 9k.

let me know if it cannot fit after 10k iters and i will double check the code.

[JunnYu]

i use sd1-5 and reproduce the resut. Thx!

[yukang123]

is that sd1 or sd2? sd1 will fit condition at about 4k to 6k iter. sd2 is usually at 6k to 9k.

let me know if it cannot fit after 10k iters and i will double check the code.

Is this rule of thumb also appliable to other training experiments on other control conditions like segmentation maps? @lllyasviel

By the way, during training the ControlNet on my own dataset, the training loss curves (accumulating the previous steps) keep fluctuating while the training learning rate is set as 10e-5. Is this normal? Or the training steps are not sufficient yet (2ksteps)？ Or the dataset size is too small? (only 6k images)

[hxy-123-coder]

@yukang123 Hello！I wonder that could you share your own dataset or where to get a dataset?

[geroldmeisinger]

I wrote a long article about control net training where I discuss the control-ini and sudden convergence https://civitai.com/articles/2078#heading-3270 . In the QA I also collect information for different step sizes.