abbuibuibui/kaguya-Lora

Kaguya Shinomiya (Kaguya-sama) — Illustrious SDXL Character LoRA

中文说明 | English (current)

An unofficial character LoRA for Kaguya Shinomiya from Kaguya-sama: Love is War, trained on waiIllustriousSDXL v17 (commonly saved as waiIllustriousSDXL_v170). This repository includes the E1–E7 epoch weights. Comparative evaluation recommends E5 as the default, at strength 0.8. The training images and captions are in abbuibuibui/kaguya-Dataset.

This is a fan-made derivative, not an official product. Rights to the character and source work remain with Aka Akasaka / Shueisha.

Files

Step checkpoints saved during training are not included. Only the epoch files are published.

Epoch	File	Role
E1	kaguya_shinomiya-000001.safetensors	Alternate. Identity is recognizable; school full-body and profile more often show a light sailor top instead of the black dress.
E2	kaguya_shinomiya-000002.safetensors	Alternate. Not advanced to the strength matrix.
E3	kaguya_shinomiya-000003.safetensors	Alternate. Screened out: the portrait picked up a hard light slash.
E4	kaguya_shinomiya-000004.safetensors	Alternate. Not advanced to the strength matrix.
E5	kaguya_shinomiya-000005.safetensors	Recommended default. Cleanest portrait and the black-dress school uniform at 0.8.
E6	kaguya_shinomiya-000006.safetensors	Alternate. Not advanced to the strength matrix.
E7	kaguya_shinomiya.safetensors	Alternate. Final-epoch filename has no -000007. Close to E5. From mid strength a red ribbon sits behind the hair more often. 1.0 overfits.

Download E5 for general use. The other six files are kept so other developers can compare epochs.

E5 SHA-256: E308F5D236A57DF3ADDF573AAE543601F99F10C220D100A488A44600B0A43C26

Showcase

These images are direct E5 @ 0.8 outputs with no retouching. Settings: Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216; Hires and ADetailer were off.

Portrait	School uniform

Casual token	No outfit token

The standing casual shot still reads as the black school dress. The night shot keeps the black dress, white collar, and red bow on a neon street; from 0.6 this prompt also puts a red ribbon behind the hair. See the notes below.

Trigger Words and Settings

Captions used keep_tokens = 2 and shuffle_caption = true. The dataset is 42 images: 34 kaguya_school_uniform, 8 kaguya_casual. Repeats 3, 7 epochs, 882 steps. Network dim 32, alpha 16. AdamW 8-bit, UNet plus both text encoders, bf16, SDPA. The saved log does not print the learning rate, so this card does not state one.

Role	Token	Typical look
Identity	kaguya_shinomiya	Black hair, red eyes, long hair
Outfit	kaguya_school_uniform	Black dress, white collar, red ribbon
Outfit	kaguya_casual	Weak on a standing full-body. Sitting shots do change into a white shirt. Only 8 training images.

Recommended: waiIllustriousSDXL v17 or an Illustrious / NoobAI-family model; LoRA strength 0.8 (0.6 already identifies the character; 1.0 turns faces aside and locks the hands); Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216.

Example:

<lora:kaguya_shinomiya-000005:0.8>,
masterpiece, best quality, very aesthetic, absurdres, anime screencap,
1girl, solo, kaguya_shinomiya, kaguya_school_uniform,
black hair, red eyes, long hair,
white collar, red ribbon,
upper body, looking at viewer

Negative prompt:

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit,
fewer digits, cropped, worst quality, low quality, signature, watermark,
username, blurry, extra person, two girls, multiple girls,
blonde hair, blue eyes

Why E5

E5 was not chosen because it was a later epoch. Round 1 screened E1–E7 at 0.8 on a portrait and a school full-body (14 images). E3’s portrait had a hard highlight. E1’s school outfit kept light sleeves. E7 added red hair ornaments. Only E1 / E5 / E7 advanced.

The strength matrix then generated 9 scenes at 0.4 / 0.6 / 0.8 / 1.0 (108 images): front, three-quarter, profile, school full-body, casual full-body, sitting, night street, no clothing words, and no trigger. E5 at 0.8 is the balance of a symmetrical face and the black dress. E1 is the light-sleeve alternate. E7 stays close to E5 and grows the red ribbon behind the hair sooner.

The full Chinese evaluation write-up is in evaluation/横向测评记录.md.

Public Horizontal Evaluation

Each sheet below is a real comparison from one scenario: columns are epochs (E1 / E5 / E7); rows are LoRA strengths (0.4 / 0.6 / 0.8 / 1.0). Images are original outputs scaled into labeled grids; nothing was retouched or replaced.

Round 1: E1–E7 @ 0.8

Front portrait	School full body

Identity and viewing angle

Front portrait	Three-quarter view

At 0.4 the face is already black hair and red eyes, with more of the collar in frame. E5 at 0.8 is the clean symmetrical close-up. At 1.0, E5 / E7 pick up a red ribbon and the face starts to turn. The three-quarter prompt mostly stays near-frontal; from 0.8, E5 / E7 add a red ribbon behind the hair. Profile holds the black dress on E5 / E7 through 0.8. E1’s profile at 0.8 drops into a light sailor top. At 1.0 the profile tilts up and the red ribbon returns.

Full body and outfit control

School full body	Casual full body

School full-body at 0.8 is the black long-sleeve dress, white collar, and red bow on E5. E1 more often keeps light sleeves. Standing kaguya_casual stays close to that black dress at 0.6 and 0.8, so the token is not a reliable outfit switch. Sitting with the same token does change into a white shirt; E1 keeps a red hair bow there, and E5 / E7 drop it around 0.8. At 1.0 the sitting pose opens up.

Scene generalization

Night street, school uniform, neon. At 0.4 all three candidates stay front-facing in the black dress. From 0.6, E5 / E7 keep the dress and add a red ribbon behind the hair, usually with a hand near the cheek. E1 at 0.8 and 1.0 leaves the black dress and turns away in a light outfit. Identity holds on E5 / E7 in this scene; the red ribbon is the failure mode, not a loss of the face.

Default-outfit leak

With kaguya_shinomiya and no clothing words, the white collar is already there at 0.4. From 0.6, E5 / E7 add the red ribbon behind the hair, and at 0.8–1.0 the face turns. Write other clothes explicitly.

Trigger isolation

The LoRA is loaded, but the prompt has no kaguya_shinomiya. It still asks for black hair, red eyes, and a school uniform, so those words explain the hair and eye color. The black dress does not lock in. At 0.4–0.6 the results are ordinary blazers or sailor collars. At 0.8–1.0, E5 / E7 drift toward a white sailor with a red bow, which is outfit-color bleed, not the trained black dress.

Training

• Data: 42 manually selected stills with matching .txt captions. Captions start with kaguya_shinomiya, <outfit>, keep_tokens = 2, remaining tags shuffled. Outfit counts: school 34 / casual 8. Published as abbuibuibui/kaguya-Dataset.
• Network: LoRA rank 32 / alpha 16, UNet plus both text encoders.
• Optimizer: AdamW 8-bit. The training log does not print UNet or text-encoder learning rates.
• Precision and resolution: bf16, SDPA, 1024 buckets, batch size 1, latents cached to disk. Flip and color augmentation off.
• Steps: repeat 3 per image, 7 epochs, 882 optimizer steps (126 per epoch). One weight file saved per epoch. Mid-epoch step files are not published.

Compatibility and Limits

This release was scored only on waiIllustriousSDXL v170, Euler a, 832×1216.

Base model	Result
waiIllustriousSDXL v170	Measured. This is the training checkpoint.
Other waiIllustriousSDXL / Illustrious-XL / NoobAI-XL merges	Same family. Not re-scored for this character. Try 0.7–0.9.
Pony Diffusion V6 XL	Not tested here. Identity can drift on this family.
Generic SDXL 1.0	Not tested here. Poor on the earlier character LoRAs in this series.
SD 1.5 / SD 2.x	Incompatible.

This is a character LoRA, not a general style model. kaguya_casual does not reliably replace the black dress on a standing shot. Three-quarter, night, and no-clothing prompts from 0.6 can add a red ribbon behind the hair. Write other clothes explicitly.

License

• Character and source-work rights remain with © Aka Akasaka / Shueisha / the relevant production committees.
• The LoRA weights are released under CreativeML OpenRAIL-M.
• Intended for non-commercial fan use. Generated images must also follow the license of the base model in use.
• Do not use this model for unlawful, deceptive, harassing, or exploitative purposes.