>>106569194
>>106569219
>>106569381
Understand that my "skill" with image editors begin and end at kindergarten doodles on MS Paint, so I really hope inpainting's all I need. Granted, what I gen now isn't bad, but I'm certain it can be better. Especially when it comes to the background. Videos are a never ever for me so the next best thing is making stuff that's indistinguishable from an anime screenshot, including the old ones. Some genners on Pixiv figured it out and I want to as well.

Also, I've taken to setting the hires multiplier to no more than 1.5 (for a typical 1248x1824) if I'm starting at one of the basic resolutions dictated by pic related, in this case 832x1216 (and its wide inversion) until further notice. Does ~1536 refer to the width, height, or both?

>>937989022
you use the sdxl guide?

>>105754461
I'm using these, I set up max resolution at 1152x1344, lycoris, network and convolution at 32 (normal and alpha) lr 0.0003, TE 0.00015, 20 epochs, 5 repeats in dataset, Huber exponential loss, if you need more parameters please let me know.

If I'm able to replicate the artsyle, tag concepts aren't correctly learned, and If I get the concepts, I get that low resolution thing.

>>101725046
XL models have specific resolution that it was trained to be able to be genned at (picrel) outside these work but more likely to give you wonky results. Also XL in general can't go straight to 1920 you have to do a two step process (What hi-res steps are, if you seen that thrown around). So you initially gen at these lower res and then increase it with the hires.