thibaud

/

controlnet-sd21

like 373

The dataset is roughly 300K images.
It takes 20-22 hours for this version.

where can I get that size of training dataset?

I use a subset of LAION

I use a subset of LAION

thanks.

should I follow the doc here: https://github.com/lllyasviel/ControlNet/blob/main/docs/train.md on how to train my sd2.1 controlNet ckpt?

Yes.

Sorry to bother you again.

There is some point that I don't understand.

For like:
If I want to build a ckpt with openPose. How I suppose to make the trained modal using the annotator openPose?

or I just run tutorial_train_sd21.py, it will automatically build all the annotator models for me?

no you have to create your own dataset and annotator dataset. Good luck. It's very time consuming!

no you have to create your own dataset and annotator dataset. Good luck. It's very time consuming!

So I need to first trained the openPose annotator dataset by combine subset of LAION + openPose

after that, I use the openPose annotator dataset to feed ControlNet? right?

thanks for sharing your models with us. It's very helpful.

haha, I get it now. I download the fill50k.zip. and see what's inside, there are 2 dirs. I suddenly understand what your own dataset and annotator dataset means. thanks.

Perfect!

Hi, does Controlnet SD2.1 need a new anotator, to preprocess the image?
I did all this,
download model, setting 15 to 21.yaml, fixed the yaml,
im using v2-1 768 noneme model.
then give a prompt and a image of city,
I was expecting SD 2-1 786 could do better job than SD1.5, I can see ControlNet is working the lineart correctly,
but the result is a crap, same promts works fine gives beautiful cityscape in SD 1.5.

sorry man, I am suing SD21 512*512, I am not sure is because of the size or not. maybe does not work for 768？

Hi, does Controlnet SD2.1 need a new anotator, to preprocess the image?
I did all this,
download model, setting 15 to 21.yaml, fixed the yaml,
im using v2-1 768 noneme model.
then give a prompt and a image of city,
I was expecting SD 2-1 786 could do better job than SD1.5, I can see ControlNet is working the lineart correctly,
but the result is a crap, same promts works fine gives beautiful cityscape in SD 1.5.

in my case any time I use any model based on SD 21 (v2.1), 512x512 gives awful result, until I change resolution to at least 768x768, I think that's probably this is what the v1.2 model was trained on. So unless you did that already, I recommand moving to at least 768px resolution in your tests.