Sango (Inuyasha: The Final Act) | Pony XL | NSFW Support

Hello! Before continuing, I would like to warn you my descriptions are long. I'm not a fan of being cryptic about information. I don't want users to have to "connect the dots".

I highly recommend that you take the time to read through everything. You'll learn the most that way. While I do enjoy kermit typewriting, I appreciate if people read everything before asking questions.

---

This LoRA is part of a Collection!

---

🔹 Model Information 🔹

The following is a Character LoRA of Sango from Inuyasha: The Final Act. Here are the things that you should know.

◽️ This LoRA went through many revisions. I'm sorry if it annoyed anyone.
◽️ This LoRA has several versions, which are explained in the section labelled "Versions"
◽️ The dataset is mostly sourced from TFA. Yashahime does not contain good character designs.
◽️ While I want my LoRA to capture her face, hair and outfit as accurately as possible - I don't really care for anime accurate body proportions or 1:1 mirrors. Does that make sense? I don't want og retro anime but I also don't want the mature woman big booba style. I aim for the middle.

---

🔹 Versions 🔹

I have multiple versions and I do believe I need to explain them in some depth..

SangoPonyV1
◽️ I originally shared the wrong version of this in the past.
◽️ Trained with AdamW8bit with 1000 steps (100 images @ 5 epochs, 10 repeats, 4 batch size)
◽️ Accidentally trained with sdxl_vae
◽️ Dataset was trained with "anime coloring, retro artstyle" style tags.

SangoPonyV2
◽️ An almost identical mirror to V1.
◽️ Trained with Prodigy with 1500 steps (100 images @ 15 epochs, 4 repeats, 4 batch size)

SangoPonyV3_a / SangoPonyV3_p
◽️ These two were a recent revision and use a different dataset, artificially increased and containing Miroku.
◽️ One version uses Adafactor, and the other uses Prodigy, which can be identified by the letter at the very end of the model name
◽️ Both were trained with 1230 steps (206 images @ 10 epochs, 2 repeats, 4 batch size)
◽️ Dataset was trained with "anime coloring, retro artstyle" style tags.

SangoPonyV4_a / SangoPonyV4_p
◽️ This version is almost a carbon copy of V3, but with some differences.
◽️ The Adafactor version was trained with 1230 steps (206 images @ 10 epochs, 2 repeats, 4 batch size)
◽️ The Prodigy version was trained with 1030 steps (206 images @ 5 epochs, 4 repeats, 4 batch size)
◽️ I removed the "anime coloring, retro artstyle" style tags from the dataset.

---

🔹 Prompt Usage 🔹

Here's everything worth of interest, or worth mentioning.

There are no outfit triggers, such as "kimono_outfit" or "bodysuit_outfit". I haven't bothered with that yet. There is only the main trigger.. "sango".

Appearance (Face)
◽️"sango, brown eyes, makeup, blunt bangs, brown hair, long hair"
◽️"sango, brown eyes, makeup, blunt bangs, brown hair, long hair, low-tied long hair"
◽️"sango, brown eyes, makeup, blunt bangs, brown hair, ponytail"
◾️ Optional tags: "black hair" (limited)

Outfit (Kimono)
◽️"kimono, pink kimono, skirt, green skirt, fingerless gloves"
◾️ Optional tags: "sandals" (limited)

Outfit (Bodysuit)
◽️"bodysuit, black bodysuit, shoulder pads, fingerless gloves"
◾️ Optional tags: "weapon, scabbard, mask, giant boomerang" (limited)

All in One (Sample Prompts)
◽️"sango, brown eyes, makeup, blunt bangs, brown hair, low-tied long hair, kimono, pink kimono, skirt, green skirt, fingerless gloves"
◽️"sango, brown eyes, makeup, blunt bangs, brown hair, ponytail, bodysuit, black bodysuit, shoulder pads, fingerless gloves"

I don't feel like mentioning the Miroku tags. Do note these are examples. Be sure to use score tags, expressions.

---

🔹 Tips and Recommendations 🔹

I'm not really a SD scientist, but I'm going to share a few points that may or may not help you.

◽️ Reference the sample images, be it the initial preview or the very NSFW extras I generate often. I do not scrub metadata, I don't do that shit. Everything is made with SD Forge
◽️ The sample images are a mix of same looking poses, and also varied x/y/z plots - it's for a reason. It's to see how well the model can perform.
◽️ All images use Prefect Pony v6.0. I don't care for SFW models, but I would say WAI and Hassaku are decent alternatives.
◽️ Use a weight between 0.8 and 1.
◽️ Use ADetailer. It doesn't work for me.