mental illness time:
https://www.mediafire.com/folder/20fjepve57u1v/Artist_Test
19,152 artists 57,456 images, e621 and danbooru.
The human and multistanding folders have a password - /msdg/
They're nsfw and some of the artists make the results look like shota so just trying to not get the account taken down.
They're all in seperate folders but if you combine them into the same folder like pic related and resize and zoom your explorer window you can view them by scrolling, pick out ones you like and copy them to another folder for later.
I've also included some scripts. One script will rename the pictures and you can use it to add a set using your model/prompt structure, just gen 9,576 or 19,152 pictures, then use the script to rename everything so it can go in the same folder again and be 4 columns instead of 3.
The other script can pull metadata from jpgs but i think you need prerequisites installed, I dunno it just works on my machine, copy the script into chat gpt and ask it how to get it running. I use it to pull the metadata then text process it to get the artist prompts out of my handpicked artists to make wildcards out of. This is what I got out of my last go at this with zoinksnoob
https://files.catbox.moe/zrxyhs.txt only about 20% of the artists do something, and they're mostly all fucked up in stablemond. This set will probably likewise be of limited value to models other than stablemond
I can't program, I just got chat gpt to make me these things so they might be literal viruses, but they work for me.
This time i didn't make the mistake of genning them as full size jpgs so I can share them without buying a Microsoft 365 sub. So that's why they're so crunchy.
I kinda fucked it up anyway but this is already gonna be a chonker without getting into that.
I chose 832x832 for speed of course, 1024 would have been better at identifying textures but it would have taken me over 140 hours even with the 15 steps from this
>>76603. Also I had an issue where the gen would fuck up randomly and only create black squares until i reloaded the model. not sure if it was the 15 step lora doing that or if my gpu's dying or something.
I did some testing and the 1024 was better but for differentiating artists 832 seemed good enough
832 -
https://files.catbox.moe/0yzyth.jpg https://files.catbox.moe/vnkunt.jpg
1024 -
https://files.catbox.moe/db0osb.jpg https://files.catbox.moe/xq4kyd.jpg
Main thing, it's looking like the zoinksnoob set i did, there's a lot of false positives in this.
Take this guy for example:
https://files.catbox.moe/d2y6zc.png
I ran it with a different prompt structure, weak lora and another artist tag, then turned the prompt up to 1, 1.5, 2 and 3
https://files.catbox.moe/oxev3o.jpg
nothing
but this guys is probably real
https://files.catbox.moe/5st0s0.png
because you can kinda see a resemblence to his e621 stuff
https://e621.net/posts?tags=salanchu
And running it through the same test produces a (slightly) noticeable and consistent effect
https://files.catbox.moe/lp3pbm.jpg
This is a fairly weak response, a kinda similar artist with more tagged images produces this effect
https://files.catbox.moe/4hj8mn.png to
https://files.catbox.moe/n6dspr.jpg
note it doesn't need to be turned up much to produce a noticeable effect over my lora and artist tag and it gets overcooked pretty fast.
Anyway there should be less useless artist prompts earlier on the square brackets number will tell you how many images there were in the training data for that artist based on the spreadsheets the noob guy's provided.
There's some weird anomalies though. like this guy
https://files.catbox.moe/hc50qz.PNG. That artist tag is not producing that effect and for some reason there are not 198 pictures on their e621
https://e621.net/posts?tags=racerdragon. Dunno, maybe they were deleted. I've found a lot of artists that produce no effect despite having enough training data are either pony or sonic artists. I think the noob trainers cut a lot more of that kind of picture than any other. not all of it or it would be able to do pony, i guess it was a balanced amount, or at least that's the impression i've gotten.
I have no idea if any of this makes sense or if anyone can even be bothered to read all that. That's quite a hefty wall of text. Save it quick if you want it. I'm probably gonna find more problems with it, get depressed and delete it sooner or later.