Senko-san (Sewayaki Kitsune no Senko-san) — Illustrious-XL LoRAHigh-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 (<code>txt2img</code>) and seamless <code>img2img</code> 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<ul><li>Identity Accuracy: Faithfully captures Senko's distinct facial structure, amber eyes, hair clip/ornament, fox ears, and cel-shaded anime aesthetic.</li><li>Inpainting &amp; Face Swaps: Excels at <code>img2img</code> inpainting. Smoothly adapts Senko's head and features onto generic character bases while respecting the target image's lighting, outfit, and background geometry.</li><li>Outfit Mutability: Learns her character identity without locking her rigidly into her default shrine maiden outfit.</li></ul>Recommended Generation Settings (For Everything)<ul><li>Base Model: Illustrious-XL</li><li>Trigger Word: <code>senko_san</code></li><li>Optional but recommended tags: 1girl, solo</li><li>LoRA Weight: <code>0.7</code> – <code>1.0</code> (Use ~0.7 for <code>txt2img</code> and <code>0.85</code> – <code>1.00</code> for <code>img2img</code> for optimal results)</li><li>Inpainting Denoising Strength: <code>0.55</code> – <code>0.7</code> (<code>0.55</code> should be your default)</li><li>Sampler: Euler (Euler provides clearer, higher quality results - use Euler Ancestral for dreamy results)</li><li>CFG Scale: 7 (works well at 4 for certain images, probably works well for inpainting with masking at 10)</li><li>Steps: 40</li><li>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)</li></ul>Recommendations for Image 2 Image (img2img):<ul><li>Ensure your input image is clear and crisp if that is the kind of output image you want to create</li><li>Use an input image of 1024 x 1024 for the highest probability of success<ul><li>Alternate resolutions can work fine as long as the image is high quality</li></ul></li></ul>Training Parameters<ul><li>Engine &amp; Model Infrastructure:</li><li>Engine: Kohya (<code>sd-scripts</code>)</li><li>Base Model: Illustrious-XL</li><li>LoRA Type: Standard <code>lora</code></li><li>Training Schedule &amp; Duration:</li><li>Target Steps: 1,475 total steps</li><li>Max Epochs: 10 epochs</li><li>Dataset Repeats: 5 repeats</li><li>Batch Size: 4</li><li>Network Architecture:</li><li>Network Rank (Dim): 32</li><li>Network Alpha: 16 (scaling factor \alpha/r = 0.5)</li><li>Optimizer &amp; Learning Rates:</li><li>Optimizer: Prodigy</li><li>UNet Learning Rate: 1.0 (standard dynamic baseline for Prodigy)</li><li>Text Encoder Learning Rate: 1.0 (standard dynamic baseline for Prodigy)</li><li>Scheduler: <code>cosine</code></li><li>Scheduler Cycles: 1 full cycle</li><li>Dataset &amp; Resolution Handling:</li><li>Base Resolution: 1024 \times 1024</li><li>Dynamic Bucketing: Enabled (<code>enableBucket: true</code>)</li><li>Shuffle Captions: Enabled (<code>shuffleCaption: true</code>)</li><li>Keep Tokens: 1 (preserves the primary trigger word)</li><li>Flip Augmentation: Disabled (<code>flipAugmentation: false</code>)</li><li>Conditioning &amp; Regularization:</li><li>Noise Offset: 0.1 (enhances shadow depth and contrast)</li><li>Min SNR Gamma: 5.0 (stabilizes loss on low-noise steps)</li><li>CLIP Skip: 1</li></ul>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: 1Parameter 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 2Example 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: 1Example 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: 1Example 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

Senko-san-Illustrious-XL-v.01

<ul><li>Trigger Word: <code>senko_san</code></li><li>Optional but recommended tags: 1girl, solo</li><li>LoRA Weight: <code>0.7</code> – <code>1.0</code> (Use ~0.7 for <code>txt2img</code> and <code>0.85</code> – <code>1.00</code> for <code>img2img</code> for optimal results)</li><li>Inpainting Denoising Strength: <code>0.55</code> – <code>0.7</code> (<code>0.55</code> should be your default)</li><li>Sampler: Euler (Euler provides clearer, higher quality results - use Euler Ancestral for dreamy results)</li><li>CFG Scale: 7 (works well at 4 for certain images, probably works well for inpainting with masking at 10)</li><li>Steps: 40</li><li>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)</li></ul>

Files with SHA256: f171b8a2275049d8e182fc2a6bf470fb4453b042ff139d828d9fa0ee3fa4e3e1