Tekuho Ultimate - v3

Finally... after 10,000 years.

Shoutout to Tekuho, who this Lora is inspired by! Please, check out his awesome work, and shower him with support!
https://www.patreon.com/user?u=6714576
These loras aren't meant to replace or compete with the OG artist, they're mainly just a way of spicing up outputs, for fun, and my own learning experience.

How to use

"tekuult" in the prompt as the trigger word, throw that bad boy in at 1.0 strength. Ez pz.
The lora by default produces a more "realistic" look, while taking cues from how Teku girls are drawn. If you're interested in the more graphic, flat look, simply add (flat shading:1.2) to the prompt! It's not a bug, it's a feature!

Works great with character loras! The style might not come through on some select loras, due to how they trained them. Most good quality loras will work great with this style lora.
Check out the showcase for some examples using char loras.

The trigger word "tekuult" isn't strictly necessary, but it's just for peace of mind.
All showcase images are genned making use of the ADetailer plugin, but results will be fine without it.

The model is trained off of the NAI checkpoint so it should work great with it and any mixes it's present in (which should cover a good number of anime checkpoints).

The nitty-gritty

What's new? Why v3? Where's v2? What's next?

What's new in v3? Glad you asked. We've got.....

• A bigger dataset! Got double the data in v3. A bigger dataset that I've once again went through with a fine-tooth comb, tagging, painting out text, fixing issues, and all that jazz. I can see Teku girls when I close my eyes now... Someone please send help.

• Higher resolution! Now that I've got a brand spankin new RTX 4070ti, v3 is trained at 768x768 resolution instead of the default of 512x512. Results are cleaner, fleshbeasts are rarer, and coherency is off the charts (compared to v1 anyway).

• More control! As I've meticulously tagged every image yet again, it should be much easier to prompt with v3. On the same token, "simple background" should rear its ugly head much less often now, as I've gone in and added new backgrounds to a large chunk of the dataset that had "simple background, white background".

• More manners! v3 was raised with good parenting in mind! As such, v3 should play real nicely with most character loras you invite over to play with it. Check out the showcase! Some Loras might not work too well and the style may not come through as strongly, but it's the fault of how the other person raised their lora, not mine! Train it right and it'll treat you right!

• More toggles! Going from 0 to 1 is still more! v3 has the feature to adjust the style of the lora! As mentioned, by default, the lora goes for a "hi-res" look, with smooth shading while still keeping the overall feel of Teku's girls. While that's all well and good, it's not exactly "The Style". I've added (read: total quirk of the model) the ability to swap between the hi-res look and the more graphic and traditional look of Teku's girls, simply by using the tag (flat shading:1.2). This will change your gens slightly in terms of composition, but the feature is neat. Have your cake and eat it too! Check out the showcase for examples of the tag in action!

• Less bleed! With all that extra VRAM, not only was I able to train at higher res, I was also able to up the batch size! Apparently, upping the batch size and training on multiple images at once lessens the effect of traits from the source image making guest appearances in your gens.

• 50% more Tammy! Literally! I swear Tammy probably makes up about 25% of the dataset now. Perfect for all those Tammy lovers. While she's not explicitly tagged by name, you can coax her out with the tags "brown hair, long hair, aqua eyes, v-shaped eyebrows". Braces aren't included in this model, as I've edited them all out of the dataset to make faces gen right, but she's still chilling in the model! Dedicated Tammy+Braces lora possibly coming in the future???

Why v3? Where's v2?
Everyone knows that 3 comes after 1!
But seriously, I trained a v2 a good while back, and I've been using it for quite a few months. That being said, I felt that the quality of it only marginally beat out v1, so I opted to not upload it. I think around that time I had also taken down v1 as well. 1. because I didn't think it was entirely fair to Tekuho (even though he's apparently fine with it going by the comments on this model). Now that the cat's out of the bag and there's a few other loras on here going for Teku's style, I figured I'd at least try to do him justice. And 2. because I wasn't really happy with the quality of it. As the v1 description used to say, "Ultimate kinda implies it will be the best, right?". I'm sure a good number of people thought that it was, but to me it wasn't to my standards. Totally not because I had trouble genning my own waifus in v1 and v2 so I made v3!

What's next?
I've been having a bunch of fun with Pony XL, it's ability to gen all of my degenerate prompts is off the fuckin charts, so maybe in the not too distant future I'll see about training a Teku Ultimate XL version on that model. I know fuck all about XL training, but we'll see!
I also have a second version of v3 trained, that includes femboys, but it requires a bit more wrangling to gen those femboys. It's all finished and ready to go, but unless there's a demand for it, I think I'll hold off on posting it until I can consistently tease em out. I had also tried to train specific Teku Girls and Boys into the model that you could call up via a tag, but the feature just didn't work. Well, except for Tammy anyway. Her tag worked. I wonder why? :^)

Anyway, that's it!

If you've read this far! Thanks! You're a real G.

Lemme know how it works for ya! I'd love to see what you come up with!