Questions about training data for poses

[AndyLonly]

I have currently completed the initial training of the SD2.1 canny model and am working on the training of Pose.

I currently have an accurate annotated training data and a large amount of unannotated data,I currently intend to use only the full body of unlabelled data to auto-labelling, and I'm wondering if this has serious implications for practical use after training? Other than that auto-labeling seems to be less accurate, sometimes even a mess, and I was wondering if there is a high percentage of this in your dataset?

less accurate:


mess:

PS: this pose estimator is for Illustrated character

[geroldmeisinger]

I collect information about "promptless training" and "prompt dropping" here https://civitai.com/articles/2078#heading-3617 -> Q: Why does a control net need captions or why would you drop captions?

you might also consider using automated caption generation like BLIP. there might even be specialized ones for characters https://huggingface.co/docs/transformers/model_doc/blip-2