Several Variants. NSFW Support. Long Description. Gallery Contains Prompt Guidance.

This is my first character Lora. I've spent over 80,000 buzz over a 8 month period, due to curiosity and interest. I am sharing the results that weren't a total failure - to the community.

If you like my work and would like to leave a review, feel free to. I do this for fun, not as a professional hobby, etc.

Last Update: 5/27/2025

General Lora Information

Inuyasha: The Final Act is the main source of my dataset(s), because Yashahime butchered the character design. Datasets are generally a little over 150 images on average. Datasets vary from variant, and more than half are lost. If they are not lost, I share them.

All variants have NSFW support. Mostly trained with Pony Checkpoints. I train them purely with Civitai as I do not have the GPU to do it locally. They are also probably overtrained in some ways. Because I'm horny I mean very curious, I typically generate excess images, feel free to use them as guidance.

Generation Tips

• Lora Strength: 0.5 to 1

• CFG: 7 to 10

• Sampler: DPM++ 2M SDE, DPM++ 3M SDE & Euler A (I use Euler A as baseline)

• Sampling Steps: 20 to 40

• Upscaler: 4x Anime generally

• Resolution: It's trained for 1024, so keep that in mind

• Other Settings: I dunno

Trigger Words/Prompt Usage (the ones that matter)

Trigger Word: "sango"
Face/Head: "brown eyes, brown hair, black hair, makeup, bangs"
Hairstyles: "long hair" or "long hair, low-tied long hair" or "ponytail"
Clothing (Kimono): "kimono, skirt, green skirt" (later versions: "pink kimono")
Clothing (Bodysuit): "armor, shoulder pads" (later versions: "black bodysuit")
Style (?): "anime coloring, retro artstyle" (you are free to remove them)
Expressions: "annoyed, angry, sad, tears, wide eyed, clenched teeth, teeth"
Extras: "mask", "gloves, fingerless gloves", "weapon, sword", "giant boomerang" (these were last minute things or weren't a main focus)

Here are two examples. First one is the Kimono, second is the Bodysuit.

Sango, 1girl, solo, brown eyes, brown hair, black hair, long hair, low-tied long hair, makeup, bangs, kimono, green skirt, skirt, anime coloring, retro artstyle

Sango, 1girl, solo, brown eyes, brown hair, black hair, ponytail, makeup, bangs, black bodysuit, shoulder pads, anime coloring, retro artstyle

They are examples. Be sure to use score tags, body tags, expressions, etc. Be sure to have decent negatives. You can probably use my negative tags in the gallery. Feel free to experiment regardless.

Her hair color, breast size, eye color, hair style can be changed afaik. Her hiraikotsu wasn't really trained or a main focus.

See below for variant-specific prompt usage.

Variant Information

Sango V2.5 - Prompt Usage & Tips

• 8225 steps (188 images @ 25 epochs, 7 repeats, 4 batch size)

• Old (6+ months ago)

• Dataset lost, tagging quality not that great

• Forgot to train her Hiraikotsu

• Trained Model: Deep Dark Hentai Mix - NSFW Anime (v6.0)

• Potentially one of the better looking faces, but prob because it's overtrained

• Overtraining resulted in sauna/hot spring/onsen/shower stuff appearing too dark. I shouldn't have included these tbh

Sango V3 - Prompt Usage & Tips

• 7480 steps (187 images @ 20 epochs, 8 repeats, 4 batch size)

• Old (6+ months ago)

• Dataset lost

• Very slight alternate to V2.5 (identical, but trained on a different model at slightly less steps)

• Trained Model: Deep Dark Hentai Mix - NSFW Anime (v6.1)
  

Final Notes:

Extras, expressions and what not have not changed. Also, V2.5 and V3 both suffer in the "make a decent looking bathing scene" department. Too dark.

Sango V4 - Prompt Usage & Tips

• Super experimental variant where I took three specific doujins and tagged/trained them within the rest of the dataset

• Shares original dataset with V2.5 and V3 (excluding the 61 images)

• 9300 steps (248 images @ 25 epochs, 6 repeats, 4 batch size) or 187 images + 61 additional images that were manually tagged doujin pages

• Trained Model: Deep Dark Hentai Mix - NSFW Anime (v6.0)

Others / Extras:

I find it is important to note how to trigger these.. "styles". By the way, they might be overtrained or bleed. It is experimental after all. There are three trigger words for these, and next to them are the full doujin names.

• "doujinmuku" - ((C60) [Toko-ya (HEIZO, Kitoen)] Muku no Chi o Nagasu Ude)

• "doujinsen" - ((C76) [Holiday School (Chikaya)] Iromatsuyoibana)

• "doujinapple" - ((C70) [Applesauce (Sada Ko-ji)] Miroku Sango + 1 | 3635+1)

To get advanced, you can mess around by adding "_page#" to the tags, but I don't recommend. Lastly, make use of "comic, grayscale, monochrome, english text, multiple views, etc"

Sango V3Rev2 - Prompt Usage & Tips

This marks the point in which I started to revise things and make improvements (or at least try to). The dataset was cut down from 178 down to 148. Backgrounds were removed and replaced with alternating simple backgrounds. Tags were improved. The issue with dark bath scenes should be fixed.

This base is about ~3 months old.

• 4440 steps (148 images @ 20 epochs, 6 repeats, 4 batch size)

• Trained Model: Deep Dark Hentai Mix - NSFW Anime (v6.1)

Others / Extras:

Not much have changed in terms of prompts, unfortunately.

Outfit Differences: "black bodysuit, shoulder pads" for the armor outfit

Weapon Differences: "giant boomerang" for her hiraikotsu? Maybe? Untested

Sango V5 - Prompt Usage & Tips

This used to be the latest. I added a few more images to my dataset, making it 151. I also believe I upscaled them.

• 7550 steps (151 images @ 20 epochs, 10 repeats, 4 batch size)

• Trained Model: Prefect Pony XL v4.0

Others / Extras:

Outfit Differences: Added "pink kimono" and "black bodysuit" on top of kimono and bodysuit respectively

Hair Differences: Missing "black hair" and "low-tied long hair", uses "brown hair" and "long hair" respectively instead (not saying you can't try, though).

Sango V4Rev2 - Prompt Usage & Tips

• This is an updated version of my V4 variant, or the "doujin" variant. I want to clarify the original variant was influenced by this. I did later train them separately, but for this they were mashed together.

• Dataset is 281 total images (153 tagged character images + 128 tagged doujins), the first half is based off V3Rev2 (backgrounds mainly edited out)

• ~8000 steps (281 images @ 25 epochs, 6 repeats, 5 batch size)

• Trained Model: The Deep Dark v6.0

Below are very important tags, the doujin tags. Feel free to look them up, they're based off doujins. They can change the style (some more noticeable than others). See V4 for the other "comic" like tags if you want to use them.

Doujin Tags: "doujin_muku, doujin_tsuki, doujin_banshou, doujin_sengoku, doujin_sengoku2, doujinapple, doujinflower" (they will be explained later)

Experiment with them. Be sure to remove "anime coloring, retro artstyle" if you use them. This is a degen heavy variant so have fun with it.

Sango V6_deep60 - Prompt Usage & Tips

• Updated version of V5, one of two (the other is trained for Prefect Pony, unavailable)

• The "vanilla" variant.

• 9180 steps (153 images @ 20 epochs, 9 repeats, 3 batch size)

• Dataset is 153 images (current), dataset contains backgrounds mainly edited out, replaced with solid colors

• Trained Model: The Deep Dark v6.0

Prompt Usage should be the same as V5.