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 (<code>txt2img</code>) as well as <code>img2img</code>, inpainting, and character face/head replacement workflows.<h2 id="what-this-model-does-best">What This Model Does Best</h2><ul><li>Character identity: Reproduces Senko's facial structure, amber eyes, blonde hair, hair ornament, fox ears, tail, and anime appearance.</li><li>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.</li><li>Scene flexibility: Works in both traditional and modern environments, including Japanese interiors, kitchens, and neon city scenes.</li><li>Img2img and inpainting: Can adapt Senko's identity to an existing composition while preserving the source image's outfit, lighting, pose, and background structure.</li><li>Clean output at full strength: Weight <code>1.0</code> produced the strongest likeness without visible overbaking in validation renders.</li></ul><h2 id="recommended-generation-settings">Recommended Generation Settings</h2>Setting Recommendation Base model Illustrious XL v0.1 or a compatible Illustrious XL checkpoint Trigger word <code>senko_san</code> Helpful tags <code>1girl, solo</code> LoRA weight <code>0.85–1.0</code> Sampler Euler Ancestral (<code>Euler a</code>) Steps <code>28</code> CFG scale <code>7</code> Resolution <code>1024 × 1024</code> Clip Skip <code>2</code><h3 id="choosing-a-lora-weight">Choosing a LoRA Weight</h3><ul><li><code>0.85</code> — recommended when clothing, accessories, and scene prompts should have more influence.</li><li><code>0.90</code> — balanced character likeness and prompt flexibility.</li><li><code>1.0</code> — recommended for maximum Senko likeness.</li><li><code>0.70</code> — still usable, but produces a softer and less forceful character likeness.</li></ul>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:<pre><code>worst quality, low quality, blurry
</code></pre>Optional extended negative prompt:<pre><code>lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, artifacts, signature, watermark, username, blurry
</code></pre><h2 id="img2img-and-inpainting-recommendations">Img2img and Inpainting Recommendations</h2><ul><li>Start with a clear, high-quality source image.</li><li>A <code>1024 × 1024</code> source image gives the most predictable results, although other high-quality SDXL-compatible resolutions can work.</li><li>Start with denoising strength <code>0.55</code>.</li><li>Increase toward <code>0.7</code> when stronger identity replacement is needed.</li><li>Start at LoRA weight <code>0.85</code> and adjust toward <code>1.0</code> for stronger character likeness.</li><li>Use a focused prompt such as:</li></ul><pre><code>senko_san, 1girl, looking at viewer, smile, hair ornament
</code></pre><h2 id="example-prompts">Example Prompts</h2><h3 id="example-1-traditional-kimono-(txt2img)">Example 1 — Traditional Kimono (<code>txt2img</code>)</h3>Prompt:<pre><code>masterpiece, best quality, senko_san, 1girl, fox ears, kitsune, kimono, smiling, tea cup, traditional japanese room
</code></pre>Negative prompt:<pre><code>lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, artifacts
</code></pre>Parameter Value Workflow <code>txt2img</code> LoRA weight <code>1.0</code> Sampler Euler a Steps 28 CFG scale 7 Clip Skip 2 Resolution 1024 × 1024 Seed 12345<h3 id="example-2-leather-jacket-and-neon-city-(txt2img)">Example 2 — Leather Jacket and Neon City (<code>txt2img</code>)</h3>Prompt:<pre><code>masterpiece, best quality, senko_san, 1girl, fox ears, kitsune, black leather motorcycle jacket, blue jeans, city street, neon lights
</code></pre>Negative prompt:<pre><code>lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, artifacts
</code></pre>Parameter Value Workflow <code>txt2img</code> LoRA weight <code>0.85</code> Sampler Euler a Steps 28 CFG scale 7 Clip Skip 2 Resolution 1024 × 1024 Seed 12345<h3 id="example-3-oversized-hoodie-(txt2img)">Example 3 — Oversized Hoodie (<code>txt2img</code>)</h3>Prompt:<pre><code>masterpiece, best quality, senko_san, 1girl, fox ears, oversized white hoodie, holding coffee mug, modern kitchen, morning light
</code></pre>Negative prompt:<pre><code>lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, artifacts
</code></pre>Parameter Value Workflow <code>txt2img</code> LoRA weight <code>0.90</code> Sampler Euler a Steps 28 CFG scale 7 Clip Skip 2 Resolution 1024 × 1024 Seed 12345<h3 id="example-4-img2img-inpainting-starting-point">Example 4 — Img2img / Inpainting Starting Point</h3>Prompt:<pre><code>senko_san, 1girl, looking at viewer, smile, hair ornament
</code></pre>Negative prompt:<pre><code>lowres, bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality
</code></pre>Parameter Value Workflow <code>img2img</code> or inpainting LoRA weight <code>0.85</code> Denoising strength <code>0.55</code> Sampler Euler a Steps 28 CFG scale 7 Clip Skip 2 Suggested input size 1024 × 1024<h2 id="training-parameters">Training Parameters</h2><h3 id="engine-and-base-model">Engine and Base Model</h3>Parameter Value Training engine <code>ostris/ai-toolkit</code> Base model Illustrious XL v0.1 Architecture SDXL LoRA type Standard LoRA Trigger word <code>senko_san</code> Training precision BF16<h3 id="training-schedule">Training Schedule</h3>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<h3 id="network-architecture">Network Architecture</h3>Parameter Value Network rank/dimension 32 Network alpha 32 Effective alpha/rank scaling 1.0<h3 id="optimizer-and-learning-rates">Optimizer and Learning Rates</h3>Parameter Value Optimizer Automagic UNet learning rate <code>0.0001</code> Text encoder learning rate <code>0.00001</code> Scheduler Cosine<h3 id="dataset-and-resolution">Dataset and Resolution</h3>Parameter Value Training images 598 Caption format Danbooru-style <code>.txt</code> captions Base resolution 1024 × 1024 Caption shuffling Enabled Preserve Tag 1 Caption dropout 0.05 Latent caching Enabled<h3 id="text-encoding">Text Encoding</h3><ul><li>Clip Skip behavior: 2</li><li>Training used the penultimate CLIP hidden state appropriate for the Illustrious XL workflow.</li></ul><h2 id="validation-summary">Validation Summary</h2>The final model was tested at LoRA weights <code>0.70</code>, <code>0.85</code>, and <code>1.0</code>, across multiple seeds and outfit prompts.<ul><li><code>0.70</code>, <code>0.85</code>, and <code>1.0</code> all produced clean images.</li><li><code>0.85</code> provided strong likeness with good outfit and scene flexibility.</li><li><code>1.0</code> produced the strongest character likeness without visible overbaking.</li><li>The model retained Senko's identity across kimono, leather-jacket, and hoodie prompts.</li><li>Clean output was possible with an empty or minimal negative prompt.</li></ul>Recommended everyday range: <code>0.85–1.0</code>.

Senko-san_NAI-V4.5-Curated

Version 3.0 UpdateVersion 3.0 was retrained using CivitAI's training infrastructure with the same general training regimen and dataset.This version offers:<ul><li>Cleaner, sharper results</li><li>Better color balance</li><li>More stable faces, hands, clothing, and fine details</li><li>Improved consistency across seeds</li><li>Better performance in both <code>txt2img</code> and <code>img2img</code></li><li>Less oversaturation and overbaking at higher LoRA strengths</li></ul>The trigger remains <code>senko_san</code>, with a recommended LoRA strength of 0.85–1.0.

Senko-san-Illustrious-0_1_illustrious_3342738_epoch_10.safetensors

generated_images.zip

v3.0

v2.0

Version Update — Complete RetrainThis version is a complete rebuild of my original Senko-san Illustrious XL LoRA.<ul><li>Expanded the training dataset from 118 to 598 captioned images</li><li>Retrained on Illustrious XL v0.1 using ostris/ai-toolkit</li><li>Increased training to 2,000 steps</li><li>Updated the network to rank 32 / alpha 32</li><li>Trained both the UNet and text encoders</li><li>Improved character likeness, outfit flexibility, and scene adaptability</li><li>Produces clean results at full LoRA strength without visible overbaking</li><li>No large negative prompt is required </li></ul>The trigger word remains:<pre><code> senko_san 
</code></pre>For this version, use a LoRA weight of 0.85–1.0:<ul><li>0.85 for greater clothing and scene flexibility</li><li>1.0 for maximum character likeness </li></ul>Recommended starting settings: Euler a, 28 steps, CFG 7, Clip Skip 2, 1024×1024.

Files with SHA256: 5a9ca0a0b44c35c6527289785f83ad8f3362a96dcf22402bfdbb92fd8500be20