ligma.pro is one of the many independent Mastodon servers you can use to participate in the fediverse.
"pissed everyone off in literally record time" - Recommended by 10 out of 10 people who, for some sad reason, have a dedicated column up to watch #fediblock.

Administered by:

Server stats:

2
active users

All-bones Jones :verified:
Public
feels weird that if i go through with this zelda decomp stuff i could realistically use stable diffusion to generate good enough reference sheets for custom models
cool_boy_mew
Public
@All_bonesJones I wonder if you could get it to outright make you a new texture sheet based on an already existing one
All-bones Jones :verified:
Public
@coolboymew i could but other people already did that for me, i'd just have to supply conversion to n64
All-bones Jones :verified:
Public
@coolboymew stapling the n64 lora straight onto the facewrap lora doesn't get useful results, but if i had a few dozen n64 sheets i could train up something that does that specifically.
i experimented with fear and hunger spritesheets for that a while back. i'd call it a success because i used a miniscule dataset and insanely unfair parameters. if i chopped up the sprites into individual poses and actually provided text files to explain what the AI was looking at, i could easily make ai-generated sprites of anything and whip up a quick script in stable diffusion to turn them into sheets
All-bones Jones :verified:
Public
@coolboymew and since i'm already rambling about it, i trained up a fear and hunger artsyle lora but i'm still not satisfied with the results. i need to refine it for a stupid project i'm working on. i also realized i could trivially use img2img to make the game look like it was drawn by someone else, and that if i got the artstyle lora to my satisfaction i could reverse the process and combine that with a spritesheet lora to easily inject any character into fear and hunger

we are rapidly approaching a point where the idea guy is the only guy that matters
Himbo Techbro

@All_bonesJones @coolboymew can you recommend a good guide on making loras and properly labeling them?

the last one I read that made any sense to me said "tag anything that's *not* the thing you're trying to train" and that's really the only thing that stuck.

All-bones Jones :verified:
Public
@r000t i feel like MAYBE i should give you more than a link but really this is so fuckin easy even a monkey could do it. if you don't want to scrape gelbooru you need at least 20 (preferably more) of the thing you want to train, then you click 6 buttons, done
Himbo Techbro
Public

@All_bonesJones Don't need to use borrus, this is not for porn.

it won't overfit with a sample like this, right?

All-bones Jones :verified:
Public
@r000t pretty small sample size tbh, 20 is the bare minimum. this thing saves the epochs of the process as separate loras so if the 10th one overfits you can use the 9th, 8th, 7th, etc until you find one that's fine. whenever i accidentally overfit it's usually just fine if i use epoch 7 instead
All-bones Jones :verified:
Public
@r000t plus you can just run the lora at a lower weight, 7th epoch at half weight can fix overfitting
Himbo Techbro
Public

@All_bonesJones Alright. How many should be full-body shots, how many should be the face, how worried should I be about getting multiple expressions, and how worried should I be about only grabbing things where the subject has high contrast with the background?

or am I overthinking this?

All-bones Jones :verified:
Public
@r000t you've overthinking it, you want
1) at least 1 full body shot from front
2) at least 1 full body shot from behind
3) (optional) at least 1 full body shot from side or various other angles
4) whatever else you can get your hands on
i've been fucking around with extremely small datasets (this is off 20 images with only 1 fullbody from front and back), and i've found that 20 high quality images is usually worse than 200 mediocre ones

focusing that much on the face seems kinda pointless
Himbo Techbro
Public

@All_bonesJones The reason I'd be focusing on the face is to hopefully get it to reliably render out features such as his hourglass facepaint, his mohawk, and the two notches cut into the side of his hair, for example.

All-bones Jones :verified:
Public
@r000t getting small details right is quantity not quality