/vt/ - Virtual Youtubers

>>45912163
>1. Is there a difference between hi-res fix and upscalers? I've noticed hires contributes greatly to image quality and detail whilst upscalers do not. Am I simply using upscalers wrong?
Kind of sort of - but it's really not that much more than an img2img upscale of your image with the same prompt. If you aren't using one of the latent upscalers, I think it might actually be the same.
>2. Is there a way to check or know which resolution a model was trained on? Or should I just assume 512x512?
Assume 512x512. Most models don't have too much trouble straying from that for base gens - within reason.
>640x640
>512x768
Probably fine
>768x512
Expect fleshblobs at the edges of the image if your prompt "has room" for the AI to screw up.
>768x768
Saw one anon suggest it as base res and was surprised not to get fleshblobs...
>Using ControlNet
You can actually go far higher...I've gone 960x960 with no ill effects, caveat emptor.
Additionally, some LoRA will be kind enough to mention this as well. ie. the bkub LoRA an anon here trained a week or two ago, he mentioned specifically that it really only worked best at 512x768...or maybe it was 512x512.
>3. Could you recommend a good starting-point coomer model to practice prompting.
Most models should be fine.

>>45912746
>Lace/frilly clothing is a really good example of this.
My favorite lace/frilly/patterned clothing yet has never been touched by latent upscaling. Just SD Upscale.
https://files.catbox.moe/tgxs69.png
https://files.catbox.moe/qo78ju.png

>>45912983
Hi-Res Fix keeps the same prompt and seed, so if you're trying to emulate it, keep the same prompt

>>44706137
I must commit penance for what I've done to her
>>44694872
I love you. My ahoys are more extremely and intricately than ever before.