Questions on dataset size and pre-processing

Nobody has claimed this yet.

I am new to controlnet training and I want to train my own pose-based controlnet. There are some questions that the training guide doesn't cover.

1. What is the recommended size of dataset? I see people use sizes like 50k, 100k, 300k in the issue section. How much does the size influence the final result?

2. How to deal with the image distortion? It seems that all the training images should have consistent image shapes. But the segmentation of human figures has different image shapes. Will the distortion destroy my controlnet?

3. Should I remove the background of in pose-based controlnet training? I am not sure about its effect on BLIP's prediction and the final result.

4. Does higher training image resolution always lead to a better result?

@lllyasviel

Thanks in advance for answering my questions!

The issue names the training guide but no file, test, or entry point. It asks for guidance on dataset size, image distortion, background removal, and resolution; there is no defined code change or completion test.

Written by the indexing model from the issue text.