Senko-san-Illustrious-XL-v.01

Senko-san (Sewayaki Kitsune no Senko-san) — Illustrious-XL LoRA

High-fidelity character LoRA for Senko-san from The Helpful Fox Senko-san, specifically tuned on the Illustrious-XL architecture. Designed for both standalone character generation (txt2img) and seamless img2img face/head swaps on existing renders. Created using a large, high-quality synthetic dataset consisting of 118 images generated with the NovelAI 4.5 Curated model.

What This Model Does Best

• Identity Accuracy: Faithfully captures Senko's distinct facial structure, amber eyes, hair clip/ornament, fox ears, and cel-shaded anime aesthetic.

• Inpainting & Face Swaps: Excels at img2img inpainting. Smoothly adapts Senko's head and features onto generic character bases while respecting the target image's lighting, outfit, and background geometry.

• Outfit Mutability: Learns her character identity without locking her rigidly into her default shrine maiden outfit.

Recommended Generation Settings (For Everything)

• Base Model: Illustrious-XL

• Trigger Word: senko_san

• Optional but recommended tags: 1girl, solo

• LoRA Weight: 0.7 – 1.0 (Use ~0.7 for txt2img and 0.85 – 1.00 for img2img for optimal results)

• Inpainting Denoising Strength: 0.55 – 0.7 (0.55 should be your default)

• Sampler: Euler (Euler provides clearer, higher quality results - use Euler Ancestral for dreamy results)

• CFG Scale: 7 (works well at 4 for certain images, probably works well for inpainting with masking at 10)

• Steps: 40

• Clip Skip: 2 (yes, the training dataset settings say Clip Skip 1, which works fine in img2img, but I do not recommend doing that and you definitely cannot use Clip Skip 1 for text to image generation)

Recommendations for Image 2 Image (img2img):

• Ensure your input image is clear and crisp if that is the kind of output image you want to create

• Use an input image of 1024 x 1024 for the highest probability of success

  • Alternate resolutions can work fine as long as the image is high quality

Training Parameters

• Engine & Model Infrastructure:

• Engine: Kohya (sd-scripts)

• Base Model: Illustrious-XL

• LoRA Type: Standard lora

• Training Schedule & Duration:

• Target Steps: 1,475 total steps

• Max Epochs: 10 epochs

• Dataset Repeats: 5 repeats

• Batch Size: 4

• Network Architecture:

• Network Rank (Dim): 32

• Network Alpha: 16 (scaling factor \alpha/r = 0.5)

• Optimizer & Learning Rates:

• Optimizer: Prodigy

• UNet Learning Rate: 1.0 (standard dynamic baseline for Prodigy)

• Text Encoder Learning Rate: 1.0 (standard dynamic baseline for Prodigy)

• Scheduler: cosine

• Scheduler Cycles: 1 full cycle

• Dataset & Resolution Handling:

• Base Resolution: 1024 \times 1024

• Dynamic Bucketing: Enabled (enableBucket: true)

• Shuffle Captions: Enabled (shuffleCaption: true)

• Keep Tokens: 1 (preserves the primary trigger word)

• Flip Augmentation: Disabled (flipAugmentation: false)

• Conditioning & Regularization:

• Noise Offset: 0.1 (enhances shadow depth and contrast)

• Min SNR Gamma: 5.0 (stabilizes loss on low-noise steps)

• CLIP Skip: 1

Examples:

Example 1 - Img2Img:

Prompt: masterpiece, very aesthetic, best quality, highres, senko_san, 1girl, solo, 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, sketch
Steps: 40, workflow: img2img, priority: low, outputFormat: jpeg, ecosystem: Illustrious, Sampler: Euler, CFG scale: 7, Clip skip: 2, Seed: 1068353682, quantity: 1

Parameter Value Prompt masterpiece, very aesthetic, best quality, highres, senko_san, 1girl, solo, 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, sketch CFG scale 7 Steps 40 Sampler Euler Seed 1068353682 Clip skip 2

Example 2 - img2img:

Prompt: senko_san, looking at viewer
Negative prompt: lowres, bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality,
Steps: 40, workflow: img2img, priority: low, outputFormat: jpeg, ecosystem: Illustrious, Sampler: Euler a, CFG scale: 4, Clip skip: 2, Seed: 61169729, quantity: 1

Example 3 - txt2img:

Prompt: senko_san, 1girl, solo, hair ornament, upper body, orange hoodie, looking at viewer, smile, outdoors, bustling city street, neon lights, night, depth of field
Negative prompt: lowres, bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality,
Steps: 40, workflow: txt2img, priority: low, outputFormat: jpeg, ecosystem: Illustrious, Sampler: Euler a, CFG scale: 7, Clip skip: 2, Seed: 1811282717, enhancedCompatibility: false, quantity: 1

Example 4 - img2img:

Prompt: senko_san, looking at viewer
Negative prompt: lowres, bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality,
Steps: 40, workflow: img2img, priority: low, outputFormat: jpeg, ecosystem: Illustrious, Sampler: Euler a, CFG scale: 7, Clip skip: 1, Seed: 1539054854, quantity: 1