Yep. Thats a bit more representative. Instead of using a single image, you could get the canny map using ControlNet and generate several that match to train from. You can even clean the canny in paint for more focused training.

@ValleyPI I like the idea of a dedicated LORA for this one, but ControlNet works wonders if you can figure it out. Then you add some segmented prompts for the goal. Had to abandon the LORA, but I'm havin too much fun with this one. I'll post a few examples of what I'm talking about. :P

@ValleyPI Whats your discord would be nice if there were simple messaging. Are there discords for this?

@VicMan @ValleyPI I was looking at this thread and wanted to point out this isn't an either/or kind of thing. Using both together amplifies this effect enormously, because ControlNet gets the pose right while the Lora captures the essence of what's going on.

Also, canny is great and all, but it forces the results to track extremely closely to the original and can be weirdly distorting in some cases. If you want to allow some variation from the original it's best to use multiple layers of less strict ControlNet models (ex. OpenPose) and Loras. I'll post an example.