>>60217405description says he used
>about 22k labeled images to train about 100 epochs on a cosine scheduler with a learning rate of 5e-6HLL usually uses about 100-200k images and 10-15 epochs. If SDXL requires significantly longer training time, I see no reason to touch it.
SDXL will never be good anyway, people who made it did everything wrong. Its VAE is shit (4 channel on picrel is SDXL), 2 text encoders are shit, unet and idea of using refiner is also shit