>>74936335>did you see any improvement in speed to training te2 onlyNo, speed is about the same, even if you train only the TE, it still needs to do a pass through the unet
After trying more things, I actually think unet+embed (so-called "pivotal tuning") is the best here. But nobody uses it because it's not supported in kohya scripts
https://files.catbox.moe/6c7wz1.jpghttps://files.catbox.moe/8hk268.jpghttps://files.catbox.moe/tlbclj.jpghttps://files.catbox.moe/zdyw21.jpghttps://files.catbox.moe/30x52p.jpg 