Kemonozume (Beast Claw) [IllustriousXL]

Details

Download Files

Model description

LoRA based on anime "Kemonozume/Beast Claw" of Masaaki Yuasa's produced by Madhouse stidio. This anime series has a unique experimental visual style, interesting characters, monsters and incredibly crazy main antagonist. I admit, this LoRA cannot right reproduce the visual style and it is focused on the characters. Be careful, character descriptions will contain spoilers.

1. Toshihiko Momota

The main character is a swordsman at the Kifūken school/dojo. Kifūken members hunt Shokujinki monsters (creatures with enormous clawed hands that can transform into humans, although it's more likely that they're humans who can transform into monsters). Toshihiko is a skilled swordsman, but doesn't possess any special powers. Due to events that occurred in his childhood, he often loses his temper during battles with monsters. He falls in love with Yuka Kamitsuki and runs away from town with her.

LoRA, as expected, draws the main character well, as he has the most training images. One downside is that almost all male characters, without context, will look like Toshihiko. And without the clarifying tag "brown hair" he won't always have brown hair.

toshihikomomota, brown hair, 1boy
toshihikomomota, brown hair, 1boy, white shirt, formal, suit, necktie
toshihikomomota, brown hair, 1boy, black hakama, hakama skirt, kimono 
hawaiian shirt, tank top, scar, weapon bag, fighting stance, holding katana, ready to draw, black hair, short hair, weapon bag, black necktie, blue suit, blue kimono, brown hair, 

2. Yuka Kamitsuki

A blonde girl who is a shokujinki monster. She accidentally falls on Toshihiko on the beach while skydiving. They then begin dating; she doesn't know he's a monster hunter, and he doesn't know she's a shokujinki. Toshihiko invites Yuka to the Kifūken dojo for her father's funeral, where she is revealed to be a monster. Afterward, they escape together.

LoRA portrays Yuka well, again negatively replacing other female characters. In the training images, many images of Yuka aren't blonde; due to the lighting, her hair has been red, gray, and brown. The "blonde hair" tag is preferable; she is more associated with it and with "1girl."

yukakamitsuki, blonde hair, 1girl
yukakamitsuki, blonde hair, blue strap dress, 1girl
yukakamitsuki, short sleeves jacket, brown jacket, black dress, 1girl
short hair, pencil skirt, eyelashes, lipstick, makeup, t-shirt, handbag, shoulder bag

3. Rie Kakinoki

A girl from the Kifūken dojo. She's in love with the main character.

LoRA associates her more with the "single hair bun" tag than with her name. However, due to the strength of the "single hair bun," it's starting to appear on other characters in the scene. To fix this, you can try reducing the strength of the "single hair bun" tag and specifying that only girl has this hairstyle "girl with single hair bun", but this is guaranteed to work.

riekakinoki, single hair bun, brown hair, 1girl
track suit, track pants
grey bodysuit, elbow pads, black dress, pearl necklace, pink apron, kimoni, wavy hair, long hair

4. Kyūtarō Ōba

One of the Kifūken trio of swordsmen. Currently the head of the Oda Corporation.

LoRA associates him primarily with the "bald" tag, which is necessary to draw the character correctly. Over the course of dozens of images of Kyūtarō wearing a hood with Mickey Mouse and rabbit ears, the lore began to make him look more animalistic with each generation. And with each generation, this became more pronounced, and his ears and fur became more animal-like. Blue also became dominant in his skin tone. To correct his skin tone, you can add the negative tags "blue skin" and "gray skin." Unfortunately, the "bald" tag is very strong, making Kyūtarō almost impossible to draw alongside other male characters.

kyuutarouoba, bald, 1boy
kyuutarouoba, bald, suit, formal, pants, necktie, 1boy
kyuutarouoba, bald, white suit, black shirt, white necktie, 1boy
fat man, grey kimono, red suit, red pants, brown shirt, brown footwear, grin, rabbit ears hood,  mouse ears hood, white shirt, collared shirt, black necktie, brown gloves, labcoat, ajirogasa

5. Bon Jōji Ōba

Kyutaro Oba's son. A man of enormous stature. Despite his intimidating appearance, he's a kind and positive person.

Unfortunately, LoRA can't draw Bon correctly. With each redesign, I increased the number of images with him, but I never thought to make a separate tag for his double chin. Sorry.

bonjojioba, white suit, white pants, fedora, buttons, purple shirt, bowtie, 1boy 
giant, size difference, mohawk, brown footwear

6. Juzo Momota

Juzo Momota is the head of the Kifūken dojo/school. He's one of the Kifūken trio of swordsmen. He uses prosthetics instead of arms. Long ago, he used a secret Kifūken technique called "Kemonozume." I won't tell you what that technique is.

LoRA can only draw close-ups of Juzo Momota accurately. There are too few images of him. There is a TV series episode featuring a younger version of him in flashbacks, but he's too different.

juzomomota, old man, grey hair, thick eyebrows, wrinkled skin, kimono, 1boy
kimono, hakama, hakama shirt, chest sarashi, prosthetic arms

7. Kazuma Momota

Juzo Momota's son and Toshihiko Momota's brother. He relies on a high-tech suit in combat. He has gray hair, but as the story progresses and he transforms, his hair becomes lighter.

LoRA doesn't know about the suit or the robot. Kazuma is also a bit difficult to draw, as his brother takes up most of the space in the "1boy" tag. The "gray hair" and "long hair" tags might help.

kazumamomota, long hair, grey hair, 1boy
kazumamomota, ponytail, grey hair, lab coat, shirt, pants, 1boy
t-shirt, blonde hair, kimono, hakama, hair slicked back, blue shirt, suit, formal, grey pants

8. Gakuto Hoobari

Yuka Kamitsuki's ex-boyfriend. The leader of a small group of Shokujinki monsters, he can transform into a monster at will. One of the most powerful Shokujinki shown in the series, despite being a monster, he's an ambiguous character.

He has limited screen time and few images, and LoRA doesn't portray him very accurately; the "big nose" tag helps a lot.

gakutohoobari, big nose, 1boy
gakutohoobari, white pants, jacket, grey shirt, print shirt, 1boy
black sclera, claws, monster hand, jumpsuit, very short hair, vest

9. Shokujinki

(sorry no image, avoid R+ rating)

When creating this LoRA, the focus was on the characters in human form. Because of this, the monsters and characters half-turned into monsters turned out poorly. I didn't use a single image of Yuki with a monster arm. And only Gakuto Hoobari has a few shots of him using a monster arm. All the Shokujinki images are lumped together and grouped under this single tag.

Shokujinki monsters are drawn terribly poorly

shokujinki, claws, monster hand, sharp teeth, oversized limbs, no humans, monster, monster boy, monster girl, horns, no eyes, large mouth, fangs, blonde hair, hair over eyes, black hair, red skin, torn clothes

10. Options

In this anime, lighting often changes the mood and theme of a scene. Assigning color tags to positive or negative cues can help adjust the scene. Here's a list of key themes:

red theme, blue theme, green theme, yellow theme
dark, greyscale, sepia

Environment tags, you should probably avoid them to avoid getting ugly backgrounds.

wooden floor, sliding doors, tile wall, outdoors, wooden wall, window, big wooden door, hallway
photo background, outdoors, dojo fence
ceiling light, brown wall, yellow wall, crack wall, broken window, white ceiling, indoors, cracked floor, ceiling
train interior, hand grip, red seat, english text, vanishing point, signboard, poster (object), train station
parquet floor
railing, blue tile
nature, bush, forest, leaf, sky

11. Notes

Due to the 1,000-image limit for training LoRA on Civitai, I had to remove many images of the main characters and their facial expressions, a young Juzo Momota, supporting characters, the environment and background, and monsters. It's unfortunate, but on the other hand, if there weren't any restrictions, I wouldn't have been able to stop. So I started making LoRA with over 2,000 screenshots, which I would never have finished.

I apologize for the monkey Saru. I only have 56 images of him, but he's not human, and that's too few for LoRA to properly render him. I could duplicate, expand, crop, and extract more images of him, increasing his dataset to 90-100 images, but then I wouldn't be able to meet the 1,000-image limit. Perhaps I'll create a separate LoRA for him someday.

There aren't enough images of the kifuken uniform for LoRA to display it correctly. The tag exists, but it doesn't work.

To prevent this LoRA from receiving an R rating, some character clothing tags had to be removed from the description.

What can you tell me about the creation of this LoRA? It took about three weeks to create this LoRA, and this is the third version/interpretation. The first version had 50 separate images with backgrounds and without characters. (networkDim": 52, "networkAlpha": 26) It had 972 images, and I finished it a week ago. Overall, it was workable, but Bon Jōji Ōba and Gakuto Hoobari were turning out really badly. I also didn't like that it wasn't achieving an anime-like style.

In the second version, I tried adding the "kemonozume" style to each image (which was, of course, a bad idea). The number of images was increased to the 1000 limit. Hair color tags were also removed for the main characters, the single hair bun for Rie was removed, the "bald" tag for Kyūtarō Ōba was removed, and the hood tags were changed when Kyūtarō wears a hat with earflaps. More images were added for Bon Jōji Ōba. Some typos were corrected. The LoRA size was increased to ("networkDim": 56, "networkAlpha": 28,) to make it easier to remember. Details. As a result, after 10 epochs and 10,000 steps, LoRA was still confusing characters, and after 11-13 epochs, it had significantly overtrained, but...

In version 3, I despaired of creating a good style lore; now I need to at least create characters. Images with environments were removed. The single tag for the "kemonozume" style was removed. Character clothing tags were re-checked and made more unified. Some bugs were fixed. I restored tags for character hair color and their features, "bald," "single hair bun," and others. For Gakuto Hoobari, a clarifying tag, "big nose," was added (which had a very positive effect on its playback; it's still not perfect, but acceptable). The number of images for Bon Jōji Ōba was increased again.

In the end, I kept version 3. The distribution of images in the dataset for characters is as follows:

Toshihiko Momota - 133

Yuka Kamitsuki - 125

Rie Kakinoki - 92

KyūtarōŌba - 120

Bon JōjiŌba - 99

Juzo Momota - 37

Kazuma Momota - 81

Gakuto Hoobari - 63

Shokujinki - 68

Duo Trio - 73

Others - 65

The end. It took 72 tries to get the cover image right.

Images made by this model

No Images Found.