Not a name I'm familiar with, but have you tried using (by spaal) in a pony checkpoint and see if it works? I've found pony is usually really good at getting close to an artist style without external help, though there's a lot I still haven't tried. It may not be as strong an effect but there should be at least a little change if it does recognize.

@Kaladae Pony have 99% of artists from e621 removed, it doesnt work

@Fenrir784 Hm. My version must be modified because most of the ones I try seem to work. But then I may also just be using the ones that do.

i made a spaal lora for pony, check it out here : https://civitai.com/models/500058/pony-xl-spaal-style-lora-furrynsfw

@average3090enjoyer Thx a lot. I actually abandoned Lora training and seeing a continuation for my 1.6 lora is awesome https://civitai.com/models/231751/spaal-tihospaal-artist-style

you can just inpaint the watermark out

Its inside the training data file

ok

thats how most artist style loras work. Especially for artists that sign their stuff. Just put watermark in the neg prompt and hope for the best.

Oh, and it seems like it doesn't work for me. Tried several artists from the PDartistsV2.txt but got the same artstyle result :c

@hartovsky833 I had the same issue. Testing with the same seed and if I use any "by" it doesn't seem to make any difference. Using the listed ponyDiffusionV6XL_v6.safetensors as well.

figured out I was using the wrong model - trying to use the base pony diffusion instead of the correct one they link to here

thats what a high data set gets ya o3o, super impresive aint it? Bro went mental. Said "Lets add like 1000 artist styles" the amount of pics he had to find is insane i bet.

The description says:

Use "by <artist>", check the PDartistsV2.txt file (in "training data", also usable as wildcard)

Replace '<artist>' with the name of the artist you want to emulate closer. So if you want to emulate a guy called 'Dave', your tag would be 'by Dave'. Meaning the training data isn't used as part of the LorA but instead is an info dump for your use to inform yourself.

If you're struggling to add the LorA, then add <lora:PDXL_artist_tags_v2:#> where # is the weight you want to give the LorA in your prompt. Same as any LorA.

Still not the artist's style? Increase the weight. Did the generation mangle? Try better negatives. Does that artist only do cartoon style with watercolours? Add 3D, sketch, and other such styles that don't fit the artist to the negatives. Vice versa for the prompt if that's still not helping as much as you want. STILL not helping? Might be that you don't actually suck at prompts and the artist just wasn't trained well, either from sheer random chance or lack of material, or any other slew of reasons. Try another artist.

If you're struggling to even get to the starting line, then: If you're using the A1111 (or a fork of it) webui hosted locally on your computer, then copy JUST the .safetensors file to your webui/models/Lora folder. Restart your webui (if it was running), and then it's ready to be used. You can find it in the LorA tab of the webui and click it to add it to your current prompt. Or just copy the <lora:PDXL_artist_tags_v2:#> tag to your prompt (once again, replacing # with your desired weight).

If you think I'm patronising you (I'm not) and you already knew all this and none of it has worked for you, then you've either broken something, forgotten something, or there's gremlins in your computer- and I can't help you with any of that. Try changing clip_step and the like until you get better results. Sorry I can't help more.

@TypeTen ok I'm an idiot i figured it out. I wasn't putting <> around the artist name so i thought i was missing something i was supposed to put elsewhere. Get so used to people using <> to show a 'just fill in x here" like <this is the space you put the artist name> instead of <put the artist name between these>, i didn't think it was actually part of the tag lol

@werewolvesarehot I've just been lucky with my gens, then. lol I also understood the use of <> the same as you, so all my time writing that giant response has been wasted! Alas.

Glad you got it sorted, though! And that you came back and showed your findings! Thanks for doing so. :)

Check out my kenket lora

@LatteLeopard I don't think a LORA for one artist is the way. It kinda misses the purpose and also making the artists mad. For me, the most interesting part is to use 4-5 different artists for an image - that adding up different aspects and in the end doesn't look like a copied artist.

@nightdancer495 you can still mix it with this Lora no problem as well as any other Lora. But SDXL(pony) doesn't do great with a mix of tons of artists. I think you would be better off sticking to SD1.5 is you wanna mash together tons of artists

@LatteLeopard yeah, I feel the same regarding sticking to SD1.5 in the end - atm, I am trying out to do both

You love kenket so much you wish you could murder them and wear their skin.
"Thanks for all that wonderful training data, now you can go fuck yourself!"

What a bunch of tools.

you still keep the score tags as usual, you just add "by ___" artist tags of your choice

@Rulethony34tano I got that down, it's just things start to look strange finish with just one or two.. would it be a better idea to add more artist tags you think?

@slkuwu just add the artist tags that you need

i've seen images that only use one (and look like that one tag), and i've seen more that use several artist tags to make an artist mix

artist mixing is pretty common especially if you've been around 1.5 furry models

@Rulethony34tano I'm very sorry if I seem like a massive stupid idiot, I guess what I'm trying to say is even though no matter what I'm doing all the pictures that come out seem to be real scuffed. I'm probably just doing something wrong with the weights, I'll probably try when I'm back home from work. I'm real sorry.

@slkuwu i don't blame you, pony xl models are honestly pretty fragile and finicky especially compared to 1.5 models.

sometimes even changing or adding a single tag is enough to break the generation.

I use the score tags as normal plus negative and positive embeddings like PDXL2 etc. If you are using automatic1111/forge you can also do what I do which is to nest weight of the "by artist" prompt so it's very strong at the very beginning of generation and gradually tapers off allowing the negative embeddings and other prompt information to gradually clean up the image after basically forcing the style. When I say nest the weight I'm using these special operators:
[[[[[[(((by unknown artist))):((by unknown artist)):.1]:(by unknown artist):.2]:by unknown artist:.3]:[by unknown artist]:.4]:[[by unknown artist]]:.5]:[[[by unknown artist]]]:.6]

It looks confusing because I'm using the "[object1:object2:0.X] - first X% is object1, then swaps to object2" function nested into itself, basically what it does is start the generation for the artist weight high at ((())) strength the first 10% of steps, and then each subsequent 10% step (represented as .1 above) it swaps the prompt and decreases the weight by 10% (represented by parentheses () and brackets [] above) until after 60% it stays at minus [[[]]] weight which is the best I've found for preserving style but also leaving enough steps to clean up the image.

An alternative way to do it that might make more sense would be like this as you can see the weights for the prompt changing in numbers:
[[[[[[(by unknown artist:1.3):(by unknown artist:1.2):.1]:(by unknown artist:1.1):.2]:(by unknown artist:1.0):.3]:(by unknown artist:0.9):.4]:(by unknown artist:0.8):.5]:(by unknown artist:0.7):.6]

I know you can use an artist's style LoRA, but models are limited. Don't worry, though. I think the best and quickest way to have a multi-artist LoRA with many artists is to merge this LoRA with other artist's style models.

In my experience with illustrious models, pony loras kind of work. They work better if you're using a merge model made up with parts of pony models. In my testing at least. It is spotty.

@sigma_gang Not particularly the ideal use case, as this LoRA would be beautiful if it were trained for Illustrious. If we could use this with Illustrious or Obsession, you'd see what we mean.

@BenjiLORA I concur. I tried, and this lora doesn't play nicely with a lotta styles in Illustrious. Not saying this lora was perfect on Pony models either, but with more artists included it would be pretty great on Illustrious models.

@bevvy Hence why the implication was that retraining the LoRA with Illustrious as the base model would likely work wonders.

obviously using a LoRA that isn't trained on that base model is going to diminish a lot if not most or all of the adaptation adherence and general strength, even if they use the same architecture (i.e SDXL, Pony, Illustrious or SeaArt loras don't work, even though they're heavy finetunes on base models)

Thanks! A little dissapointed I have so many images of penises now but it'll still help

There are numerous online. We actually need specific examples for Pony.

I know this is a bit late, but this will still work with Illustrious models quite well. I've found the trick is to set the model weights to something between 0.01-0.05 strength, and clip weights to 0.75-0.85 strength.

Would be ideal if there were a dedicated Illustrious version, but unfortunately it seems nobody has decided to make one.