/vt/ - Virtual Youtubers

>>53204975
Box cause I am a massive poopy pants coward
https://files.catbox.moe/ra5nb8.png


>>53230995
Canon proportions anya
>>53206578
>All the body parts are stable, I don't see hands or heads flipping around.
The trick is to use 2 controlnet instances one using openpose/canny/whatever and the other using the new feature "reference" that just makes sure that images look similar by whatever metric.
The potential uses for this are kinda crazy you could batch openpose an image sequence and then use loopback to put the last generated image into the reference unit and keep going like that to achieve pretty incredible image cohesion across the board.
It doesnt just stop there you could layer 5 controlnet units and fine tune every part influencing the color balance and whatnot. Layering openpose units has pretty crazy untapped potential thats only gonna expand with every feature added.
If you could somehow combine this with the automatic SSIM stepping feature in shift attention or some kind of automatic detection and inpainting like ADetailer (aside from faces) it could be near perfect with 0 human intervention.
You could probably combine all of this into an automatic script but I dont know how to do any of that so I'm just here banging rocks together.