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/Iris-Lora")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Iris (KonoSuba) — Illustrious SDXL Character LoRA
中文说明 | English (current)
An unofficial character LoRA for Iris from KonoSuba: God's Blessing on This Wonderful World!, trained on waiIllustriousSDXL v17. Publishes formal E1–E7. Recommended default: E6 (iris_rE5S745-000006.safetensors) at strength 0.8. Dataset: abbuibuibui/Iris-Dataset.
Fan-made derivative; character rights remain with their owners.
Files
| Label | File | Role |
|---|---|---|
| E1 | iris-000001.safetensors |
Alternate. Early. |
| E2 | iris-000002.safetensors |
Alternate. Early. |
| E3 | iris-000003.safetensors |
Alternate. |
| E4 | iris-000004.safetensors |
Alternate. |
| E5 | iris_rE5S745-000005.safetensors |
Alternate. Matrix candidate. |
| E6 | iris_rE5S745-000006.safetensors |
Recommended default at 0.8. |
| E7 | iris_rE5S745.safetensors |
Alternate final formal epoch. |
iris.safetensors (first-round leftover, unique SHA) is not published.
E6 SHA-256: 0af40931d17842b25da086c0815fbef770a9aedbd31479be5901dea7a5d5918e
Showcase
Direct E6 @ 0.8 outputs. Euler a, 28 steps, CFG 4.5, CLIP skip 2, 832×1216.
| Portrait | Royal dress |
|---|---|
 |
 |
Trigger Words and Settings
| Role | Token |
|---|---|
| Identity | iris_konosuba (blonde hair, blue eyes, twin tails / braid, grape hair ornament) |
| Outfit | iris_konosuba_royal_dress (only outfit in this set; 40/40) |
Recommended strength 0.8. Example:
<lora:iris_rE5S745-000006:0.8>,
masterpiece, best quality, very aesthetic, absurdres, anime screencap,
1girl, solo, iris_konosuba, iris_konosuba_royal_dress,
blonde hair, blue eyes, twin tails,
full body, looking at viewer, outdoors
Why E6
Seven formal epochs screened @ 0.8 (face + royal full body). E5–E7 entered the weight matrix (10 scenes × 0.4/0.6/0.8/1.0; no alt-outfit scene — dataset has only royal dress). E6 best overall at 0.8.
Full notes: evaluation/横向测评记录.md.
Public Horizontal Evaluation
Round 1 @ 0.8
| Front portrait | Royal full body |
|---|---|
 |
 |
Round 3: E5 / E6 / E7 × weights
| Face | Three-quarter |
|---|---|
 |
 |
Training
- 40 stills, captions
iris_konosuba, iris_konosuba_royal_dress, …,keep_tokens=2, repeats=4. - LoRA dim 32 / alpha 16, UNet + text encoders; AdamW8bit; UNet LR
1e-4, TE LR5e-5; bf16, 1024 buckets. - 7 epochs planned; resume produced formal E5–E7 under
iris_rE5S745*.
Related repositories
- HF model: abbuibuibui/Iris-Lora
- HF dataset: abbuibuibui/Iris-Dataset
- ModelScope model: abbuibuibui/Iris-Lora
- ModelScope dataset: abbuibuibui/Iris-Dataset
Compatibility and Limits
Best on waiIllustriousSDXL / Illustrious family. Only one outfit token — other clothes need explicit prompts. Without outfit token the royal dress tends to leak.
License
CreativeML OpenRAIL-M for weights. Character © original rights holders. Non-commercial fan use.
- Downloads last month
- 37