Illustrious Ver
The sample images are still analyzing... give it time.
Just updated to Illustrious. Ignore the weird prompts, I was testing the TIPO extension for Forge.
PDXL Ver
It's honestly quite accurate at weight 1.2, but feel free to mess around with the weight some.
Try the LoRA in negative for some fun results too! Check below for examples.
SD 1.5 Ver
NOTE:
Large and small breasts are easy. The secret to get medium in the LoRA is to just use the "breasts" tag without any size specifiers.
Why doesn't my output look like yours?
Check the negatives in my examples. Do you have those embeds?
I used !ADetailer because it's just a set and forget quality bump.
The images were genned at 512x768 then upscaled 1.5 times using 8x_NMKD-Superscale_150000_G with denoising 0.45. You can find it with a quick Google. Note that this upscaler is non-deterministic from my testing so your images may vary slightly. (Check the bottom of this post for info on the cover image.)
Settings and Training:
Trained on Anything v3
Genned on Anything v4.5
LoRA weight 0.8
Description
FAQ
Comments (16)
Daring! I was also in the process of making a LoRA of this artist but only for personal use due to how the nature of their art conflicts with this site, so hopefully this works out for you.
Considering the alkemanubis LoRA by someone else is still up I assumed I should be fine as long as the images are ok. This art style is just so unique that I wanted to make it available to others. Also, the hands are surprisingly good in this LoRA! If it does get taken down though that is ok too. I don't want to go against the site. Also, while I'm in the top 100 the loss of my account won't be as big of a deal as losing an account like yours haha
@thesauceer Well I would hate to lose your account, I love your Zelda LoRAs. so I hope this stays up but i downloaded it just in case. while your here, do you intend to do more styles in the future?
@Idkanymore50 Style LoRAs are what I enjoy making the most so probably. I have several other old ones that I would like to share. I also intend to remake my Harada Takehito LoRA as he is one of my favorite artists and my first attempt is quite poor. Do you have an artist you wanted a LoRA for?
@thesauceer Not particularly, just wondering since this one turned out so good, like not to sound like I'm sucking up but this is probably the best working style LoRA I've seen on this site.
@thesauceer Also one more thing, would you be willing to talk about how you make your LoRAs? My latest style models haven't really been turning out well and I was ready to just give up on styles and only make characters so I'm wondering how you achieve such good results.
@Idkanymore50 Thank you! This one and my aaaa LoRA are probably my best style LoRAs? I have a couple other styles and a couple concept LoRAs that turned out well, but getting them on the site is questionable. They are completely benign, but the artists could be problematic I guess. The only reason this one was ok for me is the quality made it worth it. I'm likely going to just make a huge zip folder of all of those and upload it somewhere. Not sure where I could share that people can actually find it though.
@Idkanymore50 Ah, you'll be really disappointed in how I make them probably, but I just follow the guide made by the LoRA gui creator. These older LoRAs are set up like that. The slightly newer ones use a smaller 96x96 setup instead of 128x128. This makes all my LoRAs huge, but I have bad results with smaller dims. I use 40 repeats and only one epoch as well which I know is unusual.
Oh some other notes. I've stopped trimming my images down to a square and just leave them as is. The only editing comes from removing signatures, text, or other miscellaneous stuff. Also, for the tagging I just leave it with the default tags most of the time. I use Grabber with gelbooru and I have it set up to download the tags with the images for me. The best LoRAs I have have huge data sets. This one had 215 images, but a couple have closer to 500.
@thesauceer I already didn't trim my images and used base tags, but I should experiment with the higher dims. I typically use a range of 1 to 10 repeats and 40 epochs so I guess I'll have to try your method. I already try to use as many images as I can get but my two most recent styles only had images in the 30s. Thank you for your time and help!
@Idkanymore50 Glad that I could be useful! Good luck and I look forward to your next works. Do enjoy your break though!
@thesauceer One more thing, sorry. By 96x96, does that refer to network dim and network alpha?
@Idkanymore50 Haven't made one in a while. I'm gonna check just to be sure, but I think it was both. That's probably why I remembered it being two numbers.
Alright yeah it's both.
@thesauceer Alright thanks, hope you can find a place to share those more problematic LoRAs you mentioned, would love to see them someday.
Nice! Now we need higegepon and some other classic artists!