Sakura Kono (Horimiya) — Illustrious SDXL Character LoRA
中文说明 | English (current)
An unofficial character LoRA for Sakura Kono (河野樱) from Horimiya, trained on waiIllustriousSDXL v17. This repository includes formal E1–E7 weights. Comparative evaluation recommends E6 as the default, at strength 0.8. Training images and captions: abbuibuibui/SakuraKono-Dataset.
This is a fan-made derivative, not an official product. Rights to the character and source work remain with their respective owners.
Files
| Epoch | File | Role |
|---|---|---|
| E1 | sakura_kono-000001.safetensors |
Early. Front portrait at 0.8 often drops the glasses; eyes stay brown. |
| E2 | sakura_kono-000002.safetensors |
Glasses start sticking. Alternate. |
| E3 | sakura_kono-000003.safetensors |
Alternate. Eyes can lean olive/hazel on the portrait. |
| E4 | sakura_kono-000004.safetensors |
Alternate. |
| E5 | sakura_kono-000005.safetensors |
Matrix candidate. Usable school uniform; apron is light pink at 0.8. |
| E6 | sakura_kono-000006.safetensors |
Recommended default. |
| E7 | sakura_kono.safetensors |
Final epoch. SHA differs from E1–E6. Close on face and uniform; cooking apron drifts cream at 0.8. |
There is no sakura_kono-000007.safetensors. The end-of-training save is the unnumbered file, and it is E7.
Download E6 for general use. The other files are kept for comparison.
E6 SHA-256: 9A45EE29A7A80D1D5E10DF14A237C25A4C5E58507BBE44CC074AC834C7F692B5
Each file is 456495868 bytes.
Showcase
Direct E6 @ 0.8 outputs, no retouching. ComfyUI, Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216. Hires and ADetailer off.
| Face | School uniform full |
|---|---|
 |
 |
| Cooking apron full | Night street |
|---|---|
 |
 |
Trigger Words and Settings
Captions used keep_tokens = 2. Training captions usually keep the name trigger and the outfit token; they often omit hair color, eye color, and glasses. Those identity words below were read from the training stills and used in evaluation. Add glasses yourself — without it, mid/late epochs still usually keep the frames on school-uniform shots, but casual / no-outfit shots drop them easily.
| Role | Token | Typical look |
|---|---|---|
| Identity | sakura_kono |
long straight brown hair with blunt bangs, brown / hazel eyes, glasses |
| Outfit | sakura_kono_school_uniform |
Navy blazer, cream/grey vest, red bow, pleated skirt, black tights, brown loafers. Main eval outfit. |
| Outfit | sakura_kono_cooking_apron |
Light pink cooking apron over a light grey shirt. Alt eval outfit. |
| Outfit | sakura_kono_red_tracksuit |
Red tracksuit |
| Outfit | sakura_kono_sports_shirt |
Sports shirt |
Recommended: waiIllustriousSDXL v17 or an Illustrious / NoobAI merge; strength 0.8 for face and school uniform; Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216. For the pink apron, stay near 0.6–0.8. At 1.0 the apron often turns cream and may pick up a coffee cup.
Example:
<lora:sakura_kono-000006:0.8>, masterpiece, best quality, very aesthetic, absurdres, anime screencap, 1girl, solo, sakura_kono, sakura_kono_school_uniform, long hair, brown hair, brown eyes, glasses, looking at viewer
Negative prompt:
lowres, bad, 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, black hair, pink hair, blue hair, green hair, white hair, short hair, twintails, blue eyes, green eyes, red eyes, purple eyes
If the eyes drift green/olive, keep brown eyes in the positive and green eyes in the negative. The training stills read as brown; Illustrious sometimes pushes them toward hazel on close-ups.
Why E6
All seven epochs were screened at 0.8 on a front portrait, school-uniform full body, and cooking-apron full body. E5 / E6 / E7 went into the weight matrix (10 scenes × 4 strengths).
| Candidate | Observed result | Decision |
|---|---|---|
| E1 | Brown hair and brown eyes already, but the portrait often loses the glasses. Uniform identity is soft. | Early alternate. |
| E2–E4 | Glasses return. Face is usable; not as settled on full-body reads as the late trio. | Alternate. |
| E5 | Stable glasses on uniform. Apron at 0.8 is light pink. Night street at 0.8 already carries a blue school bag. | Matrix alternate. |
| E6 | Portrait and school-uniform full body keep the glasses and the same quiet face at 0.6–0.8. Sitting at 0.8 is hands in lap on a teal seat. Night street at 0.8 keeps bag + knee socks without breaking identity. Apron at 0.6–0.8 stays light pink. | Recommended at 0.8. |
| E7 | Face and uniform stay close to E6. Cooking apron at 0.8 drifts toward a cream apron instead of pink. | Final-epoch alternate. |
On E6, sakura_kono_cooking_apron at 0.6–0.8 is a light pink apron and clasped hands. At 1.0 the apron goes cream and a coffee cup appears — a strength effect, not an epoch-only bug.
Without an outfit token, 0.8 becomes a mauve puff-sleeve blouse and charcoal culottes; glasses usually disappear unless you write glasses. Without the name trigger, strength 0.8 still leans toward long brown hair and a cream-cardigan school look — leakage, not a prompt-only schoolgirl.
Chinese notes: evaluation/横向测评记录.md.
Public Horizontal Evaluation
Sheets use columns = epochs, rows = LoRA strengths. Original outputs only. Eval images were generated on a DGX with ComfyUI, then copied back. No retouching.
Round 1: E1–E7 @ 0.8
| Front portrait | School uniform full |
|---|---|
 |
 |
Round 3: E5 / E6 / E7 × 0.4 / 0.6 / 0.8 / 1.0
| Face | Three-quarter | Profile |
|---|---|---|
 |
 |
 |
| School uniform full | Cooking apron full | Sitting |
|---|---|---|
 |
 |
 |
| Simple bg | Complex night | No outfit |
|---|---|---|
 |
 |
 |
Training
- Base: waiIllustriousSDXL v17
- Network: LoRA rank 64, alpha 32
- Optimizer: AdamW, bf16
- Resolution: SDXL buckets, long edge cap 1536
- Epochs: 7
- Dataset: 67 image–caption pairs (School_Uniform / Cooking_Apron / Red_Tracksuit / Sports_Shirt), repeats 3
- Train host: NVIDIA DGX Spark
- Eval host: same DGX, ComfyUI (Euler a / 28 / CFG 4.5 / CLIP skip 2 / 832×1216)
Related
- Hugging Face model: https://huggingface.co/abbuibuibui/SakuraKono-Lora
- Hugging Face dataset: https://huggingface.co/datasets/abbuibuibui/SakuraKono-Dataset
- ModelScope model: https://www.modelscope.cn/models/abbuibuibui/SakuraKono-Lora
- ModelScope dataset: https://www.modelscope.cn/datasets/abbuibuibui/SakuraKono-Dataset
Compatibility and Limits
Illustrious / NoobAI lineage checkpoints work best. Other SDXL bases may keep composition but lose identity. Strength above 1.0 is usually unnecessary and makes the apron color / props drift. Glasses are part of the character; put them in the prompt for casual outfits. Without the name trigger the face still leaks at mid strengths.
License
creativeml-openrail-m
- Downloads last month
- 53