Senko-san-Illustrious-XL-v.01

High-fidelity character LoRA for Senko-san from The Helpful Fox Senko-san, trained specifically for Illustrious XL v0.1.

This version was rebuilt from a high-quality synthetic dataset of 598 tagged images generated with NovelAI 4.5 Curated. It is designed for standalone character generation (txt2img) as well as img2img, inpainting, and character face/head replacement workflows.

What This Model Does Best

• Character identity: Reproduces Senko's facial structure, amber eyes, blonde hair, hair ornament, fox ears, tail, and anime appearance.

• Outfit flexibility: Maintains her identity without forcing the default shrine maiden outfit. Validation prompts worked well with a kimono, leather jacket and jeans, and an oversized hoodie.

• Scene flexibility: Works in both traditional and modern environments, including Japanese interiors, kitchens, and neon city scenes.

• Img2img and inpainting: Can adapt Senko's identity to an existing composition while preserving the source image's outfit, lighting, pose, and background structure.

• Clean output at full strength: Weight 1.0 produced the strongest likeness without visible overbaking in validation renders.

Recommended Generation Settings

Setting Recommendation Base model Illustrious XL v0.1 or a compatible Illustrious XL checkpoint Trigger word senko_san Helpful tags 1girl, solo LoRA weight 0.85–1.0 Sampler Euler Ancestral (Euler a) Steps 28 CFG scale 7 Resolution 1024 × 1024 Clip Skip 2

Choosing a LoRA Weight

• 0.85 — recommended when clothing, accessories, and scene prompts should have more influence.

• 0.90 — balanced character likeness and prompt flexibility.

• 1.0 — recommended for maximum Senko likeness.

• 0.70 — still usable, but produces a softer and less forceful character likeness.

A large negative prompt is not required. Clean results were produced with an empty negative prompt as well as a short quality negative.

Suggested light negative prompt:

worst quality, low quality, blurry

Optional extended negative prompt:

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, artifacts, signature, watermark, username, blurry

Img2img and Inpainting Recommendations

• Start with a clear, high-quality source image.

• A 1024 × 1024 source image gives the most predictable results, although other high-quality SDXL-compatible resolutions can work.

• Start with denoising strength 0.55.

• Increase toward 0.7 when stronger identity replacement is needed.

• Start at LoRA weight 0.85 and adjust toward 1.0 for stronger character likeness.

• Use a focused prompt such as:

senko_san, 1girl, looking at viewer, smile, hair ornament

Example Prompts

Example 1 — Traditional Kimono (txt2img)

Prompt:

masterpiece, best quality, senko_san, 1girl, fox ears, kitsune, kimono, smiling, tea cup, traditional japanese room

Negative prompt:

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, artifacts

Parameter Value Workflow txt2img LoRA weight 1.0 Sampler Euler a Steps 28 CFG scale 7 Clip Skip 2 Resolution 1024 × 1024 Seed 12345

Example 2 — Leather Jacket and Neon City (txt2img)

Prompt:

masterpiece, best quality, senko_san, 1girl, fox ears, kitsune, black leather motorcycle jacket, blue jeans, city street, neon lights

Negative prompt:

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, artifacts

Parameter Value Workflow txt2img LoRA weight 0.85 Sampler Euler a Steps 28 CFG scale 7 Clip Skip 2 Resolution 1024 × 1024 Seed 12345

Example 3 — Oversized Hoodie (txt2img)

Prompt:

masterpiece, best quality, senko_san, 1girl, fox ears, oversized white hoodie, holding coffee mug, modern kitchen, morning light

Negative prompt:

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, artifacts

Parameter Value Workflow txt2img LoRA weight 0.90 Sampler Euler a Steps 28 CFG scale 7 Clip Skip 2 Resolution 1024 × 1024 Seed 12345

Example 4 — Img2img / Inpainting Starting Point

Prompt:

senko_san, 1girl, looking at viewer, smile, hair ornament

Negative prompt:

lowres, bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality

Parameter Value Workflow img2img or inpainting LoRA weight 0.85 Denoising strength 0.55 Sampler Euler a Steps 28 CFG scale 7 Clip Skip 2 Suggested input size 1024 × 1024

Training Parameters

Engine and Base Model

Parameter Value Training engine ostris/ai-toolkit Base model Illustrious XL v0.1 Architecture SDXL LoRA type Standard LoRA Trigger word senko_san Training precision BF16

Training Schedule

Parameter Value Training steps 2,000 Batch size 2 Gradient accumulation 1 UNet training Enabled Text encoder training Enabled Gradient checkpointing Enabled Checkpoint interval Every 200 steps

Network Architecture

Parameter Value Network rank/dimension 32 Network alpha 32 Effective alpha/rank scaling 1.0

Optimizer and Learning Rates

Parameter Value Optimizer Automagic UNet learning rate 0.0001 Text encoder learning rate 0.00001 Scheduler Cosine

Dataset and Resolution

Parameter Value Training images 598 Caption format Danbooru-style .txt captions Base resolution 1024 × 1024 Caption shuffling Enabled Preserve Tag 1 Caption dropout 0.05 Latent caching Enabled

Text Encoding

• Clip Skip behavior: 2

• Training used the penultimate CLIP hidden state appropriate for the Illustrious XL workflow.

Validation Summary

The final model was tested at LoRA weights 0.70, 0.85, and 1.0, across multiple seeds and outfit prompts.

• 0.70, 0.85, and 1.0 all produced clean images.

• 0.85 provided strong likeness with good outfit and scene flexibility.

• 1.0 produced the strongest character likeness without visible overbaking.

• The model retained Senko's identity across kimono, leather-jacket, and hoodie prompts.

• Clean output was possible with an empty or minimal negative prompt.

Recommended everyday range: 0.85–1.0.