Instructions to use abbuibuibui/TsubameKoyasu-Lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use abbuibuibui/TsubameKoyasu-Lora with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("waiIllustriousSDXL-v17", dtype=torch.bfloat16, device_map="cuda") pipe.load_lora_weights("abbuibuibui/TsubameKoyasu-Lora") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
Tsubame Koyasu (Kaguya-sama) — Illustrious SDXL Character LoRA
中文说明 | English (current)
An unofficial character LoRA for Tsubame Koyasu from Kaguya-sama: Love is War, trained on waiIllustriousSDXL v17 (waiIllustriousSDXL_v170). This repository publishes E1–E6 plus the unnumbered final weight. Comparative evaluation recommends E5 (tsubame_koyasu-000005.safetensors) as the default at strength 0.8.
Training images and captions: abbuibuibui/TsubameKoyasu-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 (*-step*) and optimizer state dirs are not published.
| Label | File | Role |
|---|---|---|
| E1 | tsubame_koyasu-000001.safetensors |
Alternate. |
| E2 | tsubame_koyasu-000002.safetensors |
Alternate. |
| E3 | tsubame_koyasu-000003.safetensors |
Screened into matrix. |
| E4 | tsubame_koyasu-000004.safetensors |
Alternate. |
| E5 | tsubame_koyasu-000005.safetensors |
Recommended default @ 0.8. |
| E6 | tsubame_koyasu-000006.safetensors |
Close to E5; more outfit lock at 1.0. |
| Final | tsubame_koyasu.safetensors |
Unnumbered final (evaluate vs E6 before treating as duplicate). |
E5 SHA-256: F689769DA90F6D138C25C7F691F47C7F03B35FBA9698649E68F72486099C864D
Showcase
Direct E5 @ 0.8 outputs, no retouching. Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216; Hires / ADetailer off.
Trigger Words and Settings
Captions used keep_tokens = 2, shuffle_caption = true. Dataset: 42 image/caption pairs. num_repeats = 4. Network dim 32 / alpha 16, AdamW 8-bit, UNet + both text encoders, bf16, SDPA, 1024 buckets. Training log does not print LR, so this card does not invent one.
| Role | Token | Notes |
|---|---|---|
| Identity | tsubame_koyasu |
Pink hair, pink eyes, short hair, black hair ribbon, earrings |
| Outfit | tsubame_casual |
Main (22) |
| Outfit | tsubame_hoodie |
Alternate (13) |
| Outfit | tsubame_school_uniform |
5 |
| Outfit | tsubame_tracksuit |
2 |
Recommended: waiIllustriousSDXL v17 / Illustrious family; strength 0.8. Euler a, 28, CFG 4.5, CLIP skip 2, 832×1216.
Example:
<lora:tsubame_koyasu-000005:0.8>,
masterpiece, best quality, very aesthetic, absurdres, anime screencap,
1girl, solo, tsubame_koyasu, tsubame_casual,
pink hair, pink eyes, short hair, black hair ribbon, earrings,
portrait, looking at viewer
Negative (baseline + color-bleed guards):
lowres, bad, bad anatomy, bad hands, text, error, missing fingers, extra digit,
fewer digits, cropped, worst quality, low quality, low score, bad score,
average score, signature, watermark, username, blurry, extra person, two girls,
multiple girls, blonde hair, green hair, blue eyes, green eyes,
brown hair, black hair
Why E5
Round 1 screened E1–E6 (+ Final) at 0.8. Round 3 matrix used candidates E3 / E5 / E6 across 10 scenes × 0.4/0.6/0.8/1.0. E5 was the most balanced at 0.8; E6 is close but locks outfits harder at 1.0. See evaluation/横向测评记录.md.
Public Horizontal Evaluation
Sheets: columns = candidates (E3 / E5 / E6), rows = weight 0.4–1.0. Original outputs scaled with labels only — no retouching or swapped images.
Identity / angle
Full body / outfit
Scene / leak / trigger
Training
- Data: 42 curated stills + matching
.txt. Published as abbuibuibui/TsubameKoyasu-Dataset. - Single training run:
tsubame_koyasu-000001…000006plus unnumbered final. - Network: LoRA rank 32 / alpha 16, UNet + both text encoders.
- Optimizer: AdamW 8-bit. LR not printed in logs.
- Precision: bf16, SDPA, 1024 buckets, batch 1, latents cached.
Related repositories
| Hugging Face | ModelScope | |
|---|---|---|
| Model | TsubameKoyasu-Lora | TsubameKoyasu-Lora |
| Dataset | TsubameKoyasu-Dataset | TsubameKoyasu-Dataset |
Compatibility and Limits
Scored on waiIllustriousSDXL v170. Character LoRA only. Prefer explicit outfit tokens for hoodie / school / tracksuit. High-school character.
License
- Character / source rights © Aka Akasaka / Shueisha / production committees.
- Weights under CreativeML OpenRAIL-M.
- Fan non-commercial use intended; follow the base model license for generations.
- Do not use for unlawful, deceptive, harassing, or exploitative purposes.
- Downloads last month
- 68













