Kuku (Yamainu) | Toaru Yama no Kaii ni Tsuite [Toriburi]/とある山の怪異について [トリブリ] - V1

NSFW

✨ Kuku (Yama Inu) from Toaru Yama no Kaii ni Tsuite

— She is an old inu yokai living deep in the mountains, carrying the calm presence of someone who has watched many seasons come and go. Even with that age and distance from humans, she ended up taking in an abandoned boy and raising him as her own. There is something warm and grandmotherly about the way she spoils him and takes care of every little need, but that kindness sits next to a more unsettling side. The longer she dotes on him, the more that protective love starts mixing with a dangerous instinct she can no longer keep buried.

If only I can meet grandma like her...

653868

If you know, you know~

>Doujinshi COLLECTION<

⚠️Turn off your R/X/XXX Filter to see hidden JUICY Loras in >HERE<

🖼️ How I Generate My Images

I made a full article that explains how I generate my images — including my settings, prompting tips, and even my potato PC setup 😅

👉 Check it out here if you’re curious or want to try the same method.

🔞 NSFW gallery (18+)

Check my account here to see the spicy showcase that I can't show here, if any.

🎨 Prompt Tips

These are optional starting points for handling a few behaviors of this version. You may still need some retries depending on the checkpoint, prompt, and LoRA strength.

🖤 Manga-like result

For a result closer to the original doujin’s greyscale manga appearance, avoid positive tags that strongly push anime coloring, vivid colors, or an anime screenshot style. It is also better to avoid character or outfit color tags, since they may encourage the checkpoint to colorize the image.

Recommended:

✅ Positive prompt: greyscale

📏 Reproducing her canon figure

To give the model a clearer instruction for her small build, try:

✅ Positive prompt: petite, l*li

💇 Forcing a specific hairstyle

The three hairstyles can clash with each other, especially since two of them share overlapping elements like buns or sidelocks.

If you want one specific hairstyle, it usually helps to push the others into the negative prompt and increase the emphasis on the one you actually want.

For example, if you want sidelocks_tied_back, try something like this:

✅ Positive prompt: sidelocks_tied_back
🚫 Negative prompt: long_sidelocks, ring_sidelocks

You can apply the same idea to the other hairstyles as well. If one hairstyle keeps bleeding into another, suppress the competing hairstyle tags instead of only stacking more positive tags.

💍 `ring_sidelocks` may generate an actual ring

If you prompt ring_sidelocks, it can sometimes give her a ring on her finger.

This happens because ring_sidelocks is a low-post-count tag on Danbooru, so the model may partially latch onto the word ring itself instead of understanding the hairstyle cleanly.

To reduce that behavior, try:

🚫 Negative prompt: ring

🌈 Recoloring the outfit

Since the dataset is greyscale-only, color control works best when you attach colors directly to the garment tags.

Recommended format:

✅ Positive prompt: [color]_[garment_name]

Examples:

white_yukata
red_sash
grey_shorts

This usually works better than mentioning a color separately, because the garment tag acts as the anchor. Just keep in mind that some small design parts are not cleanly taggable, so those areas may still need retries.

🎀 About the waist sash bow

If her waist sash keeps showing a bow-like back detail, that is most likely the checkpoint bias I mentioned on about version section.

At the moment, there is no guaranteed prompt fix for it. Lowering the LoRA weight or changing prompts may reduce it sometimes, but it is not something I would consider fully controllable in this version.

🧥 Default outfit details appearing on other outfits

When a default outfit detail keeps appearing in a different outfit, add the exact leaking detail to the negative prompt.

For example, if the bottom wear keep leaning to bike shorts when you do not want it:
🚫 Negative prompt: bike_shorts

Use the specific leaking tag instead of adding many unrelated negative tags.

⚖️ LoRA strength and outfit flexibility

A weight of 1 can still work with a very different outfit, especially when the new outfit is clearly different from her default one.

Only lower the LoRA weight gradually when outfit details keep leaking even after using clear outfit tags and negative prompts. Lowering it too much may also weaken her face, hair, and overall character identity.

📊 Lora Project

If you want to know what characters that I'm covering next you can check these link :
👉 Personal Project
👉 Commision Project

If this model sparked something in you, show some love with a like, review, or buzz.
Now go—

Be wild

Description

⚫ Greyscale-only dataset

This LoRA was trained using greyscale images only. Because of that, every color shown in my showcase is based on my own preferences and interpretation rather than strict source accuracy.

💇 Three hairstyle variations

This version was trained with three hairstyle patterns:
- long_sidelocks, long_hair, hair_down
- single_hair_bun, sidelocks_tied_back
- single_hair_bun, ring_sidelocks

The main problem here is tag controllability. ring_sidelocks is not a very well-known tag on Danbooru, so it can be harder to steer reliably. sidelocks_tied_back does exist, but it has a fairly small number of posts, and since it is a long tag, it may sometimes weaken into a more generic sidelocks interpretation instead of giving the exact hairstyle cleanly.

👘 Outfit pieces are individually tagged

Each garment piece was tagged separately, so you can recolor most of them however you want.

That said, some small or awkward design elements are hard to control because there simply are no clean tags for them. For example, the inner part of her yukata has a bold striped section with a different color, but I do not think there is a good tag that directly represents that specific area, so it may be harder to color that part exactly as intended.

🎀 Waist sash bias

Her waist garment was trained using waist_sash. Even though the dataset does not clearly show any back bow detail, a back-bow-like shape can still appear when she wears that outfit.

This seems to come more from checkpoint bias than from the dataset itself, and I was not really able to get rid of it. Even explicitly using back_bow logic does not reliably solve the issue.

🩳 Shorts detail limitation

Her shorts are supposed to have small dangling bow details and cutout elements at the back.

Unfortunately, those features are tiny and only visible in a small number of panels, so they were never going to train very strongly. This version is not optimized for reproducing those details reliably.

🐾 Tail tip color limitation

Her tail is supposed to have a colored tip, but that detail is not tagged in this version.

I considered using colored_tips, but I decided against it because the tag is too broad and could easily be misinterpreted during generation, such as affecting her ears as well instead of only the tail. On top of that, the colored tail tip does not appear often enough in the dataset to become a strong learned detail.

Because of that, her tail tip color may be inconsistent in generation. Sometimes it may appear with a random colored tip, and other times it may appear with no colored tip at all.

As usual, the final result can still depend on your checkpoint, prompt, LoRA strength, CFG scale, and generation settings.