Senko-san — Illustrious XL v0.1 (HD ULTRA Version)

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

This LoRA was trained on a high-quality synthetic dataset of 1,000 crisp, native 1024×1024 HD images generated with NovelAI 5.0 Curated. It produces images with a sharp, highly detailed look with precise character features and high-resolution rendering. It is designed for standalone character generation (txt2img) as well as img2img, inpainting, and character face/head replacement workflows.

This is the ULTRA version because it was trained on a dataset that includes both the original 598-image base dataset and an all-new 402-image activities dataset.

(Note: If you want the softer, standard-definition anime screencap look, check out my separate SD ANIME Senko-san LoRA (coming soon). This release is built specifically for modern, maximally sharp, high-resolution generation).

Quick Generation Settings

Setting Recommended Value Base Model Illustrious XL v0.1 (or compatible checkpoints) Trigger Tag senko_san Helpful Tags 1girl, solo, hair ornament LoRA Weight 0.85 – 1.0 (0.85 for outfit flexibility, 1.0 for full likeness) Sampler Euler a Steps 28 CFG Scale 7 Resolution 1024 × 1024 Clip Skip 2

LoRA Weight

• 0.85 — good character likeness with more room for outfit and scene control.

• 0.90 — balanced likeness and prompt flexibility.

• 1.0 — strongest Senko likeness.

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

Negative Prompt

Heavy negative prompting is not required. A simple negative prompt is sufficient:

worst quality, low quality, blurry

Optional extended negative if you encounter artifacts:

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

Example Prompts

1. Traditional Shrine Maiden / Kimono (txt2img)

Prompt:
masterpiece, best quality, senko_san, 1girl, fox ears, kitsune, kimono, smiling, tea cup, traditional japanese room
Weight: 1.0 | Steps: 28 | Sampler: Euler a | CFG: 7

2. Modern Streetwear (txt2img)

Prompt:
masterpiece, best quality, senko_san, 1girl, fox ears, kitsune, black leather motorcycle jacket, blue jeans, city street, night, neon lights
Weight: 0.85 | Steps: 28 | Sampler: Euler a | CFG: 7

3. Casual Morning (txt2img)

Prompt:
masterpiece, best quality, senko_san, 1girl, fox ears, oversized white hoodie, holding coffee mug, modern kitchen, morning light
Weight: 0.90 | Steps: 28 | Sampler: Euler a | CFG: 7

4. Img2img / Inpainting Replacement

Prompt:
senko_san, 1girl, looking at viewer, smile, hair ornament
Weight: 0.85–1.0 | Denoise: 0.55–0.70 | Resolution: 1024 × 1024

Dataset & Tagging Procedure

The dataset contains 1,000 high-resolution 1024×1024 images.

Tagging followed a multi-stage workflow combining two automated taggers, human review, custom tooling, and AI-assisted policy iteration:

1. Two-Stage Auto-Tagging:

   • Step 1: An initial auto-tagging pass was performed using Civitai's built-in auto-tagger.

   • Step 2: A subsequent pass was run using the new PixAI auto-tagger (available on Hugging Face).

2. Manual Review & Baseline Policy:

   • A representative set of 102 images was manually reviewed by hand to establish the initial tagging dictionary and baseline rules.

   • An initial draft policy was created interactively with an AI agent based on these reviewed images.

3. Custom Tooling:

   • Tag curation was managed using a customized version of toshiaki1729's dataset editor, modified to allow AI agents to assist with dataset review and tag management.

   • The tool takes the raw auto-tagger output, applies the active policy rules, and outputs the curated tag set.

4. AI-Assisted Policy Iteration:

   • The auto-tagger outputs were loaded into the custom editor.

   • An AI agent analyzed the output to surface edge cases, normalize inconsistent terminology, and categorize tags under explicit policy rules: Keep, Reject, or Unknown.

   • As new cases were surfaced, the policy was refined interactively and reapplied across the dataset to strip noise, prevent concept bleed, and produce the final curated tag set.

Training Specifications

| Parameter | Setting |

|---|---|

| Training Engine | ostris/ai-toolkit |

| Base Model | Illustrious XL v0.1 |

| Architecture | SDXL |

| Precision | BF16 |

| Training Steps | 2,000 |

| Batch Size | 2 |

| Gradient Accumulation | 1 |

| Network Rank / Alpha | Dim 32 / Alpha 32 |

| Optimizer | Automagic |

| UNet Learning Rate | 1e-4 |

| Text Encoder Learning Rate | 1e-5 |

| LR Scheduler | Cosine |

| Tag Shuffling | Enabled |

| Preserve Tag | 1 | (Preserve Tag 1: senko_san)

| Tag Dropout | 0.05 |

| Clip Skip | 2 |