High-fidelity character LoRA for Shiro-sama (シロ様) from The Helpful Fox Senko-san, trained on Illustrious XL v0.1.

This LoRA was built 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, with an emphasis on character likeness, clean linework, and outfit flexibility.

Trigger word

shiro_sama

Recommended settings

Use Illustrious XL v0.1 or a compatible checkpoint. The three examples below were generated with Nova Furry XL — IL v18.0 A, at a LoRA weight of 1.0. Start at 1.0 for stronger character influence; try 0.90 for a balance of likeness and prompt flexibility, or 0.85 when changing her outfit or giving the scene more control.

A practical starting configuration is Euler a, 50 steps, 1024 × 1024 resolution, and Clip Skip 2. The supplied examples use CFG scales of 4, 4.5, and 7. Start around 4–4.5 and adjust for your checkpoint and prompt.

For img2img, inpainting, or face/head replacement, a denoise range of 0.55–0.70 provides a starting point; adjust according to how much of the source image you want to change.

Default outfit and clothing changes

Shiro’s default outfit is her white strapless dress and red shawl ensemble: a white dress with side slits and a gold crescent motif on the bodice, a deep-red shawl draped around her upper arms, black elbow-length gloves, black thigh-high stockings, and a broad black choker with a red diamond-shaped ornament.

For the default wrap, use red shawl rather than red scarf. A miko, kimono, or hakama outfit is an alternate costume.

If shiro_sama alone does not recover her default outfit, reinforce it with white dress, strapless dress, side slit, red shawl, black gloves, elbow gloves, black thighhighs, black choker. These are generation-time garment descriptions; shiro_sama remains the character trigger. Use the baseline garment tags for the pieces you want to retain when making small changes.

When adding a jacket, omit strapless dress and explicitly describe the jacket. To keep the underlying dress and black accessories, retain white dress, side slit, black thighhighs, black choker, then add dark blue jacket or dark navy jacket. The jacket examples below use this approach.

If you want a shirt and skirt instead, describe white shirt and white skirt as the replacement clothing. Use short tail to reinforce the intended tail length.

Avoid detailed eyes if you want to preserve her black pupils; that phrase can cause them to disappear.

Negative prompting

For a minimal negative prompt, start with:

worst quality, low quality, blurry

The examples below include the longer negative prompts used for those generations.

Example 1: Nighttime skyline with a jacket and wink

Prompt

masterpiece, shiro_sama, dark blue jacket, white dress, side slit, black thighhighs, black choker, 1girl, solo, nighttime, night, city skyline, smile, short tail, wearing jacket, jacket worn, ;), wink, relaxed

Negative prompt

tears, crying, cry, sweat, rain, worst quality, low quality, lowres, jpeg artifacts, bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, extra arms, extra legs, disconnected limbs, multiple tails, detached tail, cat ears, human ears, photorealistic, 3d, cgi, realistic skin, painterly, gradient shading, monochrome, muted colors, underexposed, blurry face, multiple girls, crowd, child, loli, petite, text, watermark, signature, long tail, belt

LoRA weight: 1.0. Checkpoint: Nova Furry XL — IL v18.0 A. Steps: 50. Sampler: Euler a. CFG scale: 4. Seed: 662278959. Size: 1024 × 1024. Clip Skip: 2.

Example 2: Halloween cookie decorating in an alternate outfit

Prompt

shiro_sama, anime, close waist-up view in a vertical composition, standing behind a kitchen counter while decorating a practice batch of Halloween cookies. She leans slightly toward the viewer, her left forearm resting lightly on the counter, head gently tilted, looking directly at the viewer with relaxed brows and a small, quietly proud closed-mouth smile. Her right hand offers a ghost cookie toward the viewer at chest height, holding its lower edge so the decoration remains visible, beside rather than covering her face. The cookie has white icing, a scalloped ghost hem, two accidentally pointed upper corners, slightly uneven dark icing eyes, and a tiny foxlike smile. Her left hand loosely holds a small piping bag above parchment. She wears an ivory blouse with a rounded collar and sleeves rolled to her elbows beneath a cinnamon-orange-and-cream gingham apron with cream edging, a gathered waist, and a small pumpkin-embroidered pocket dusted with flour. Both hands remain inside the frame. The counter fills the lower quarter of the image, with a tray of decorated ghost cookies, a practice sheet of wonky icing faces, a bowl of white icing, and a folded tea towel. Behind her, a paper ghost-and-bat garland hangs beneath a shelf, small orange and ivory pumpkins sit on the windowsill beside a vase of autumn leaves, and a pumpkin-shaped cookie jar rests against cream cabinetry. Soft afternoon sunlight enters from the left, warming her face, the offered cookie, and the wooden counter. Warm cream, cinnamon orange, and muted plum accents, a cozy kitchen decorated for Halloween, her face and offered cookie the main focus.

Negative prompt

tears, crying, cry, sweat, rain, worst quality, low quality, lowres, jpeg artifacts, bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, extra arms, extra legs, disconnected limbs, multiple tails, detached tail, cat ears, human ears, miko, kimono, hakama, photorealistic, 3d, cgi, realistic skin, painterly, gradient shading, monochrome, muted colors, underexposed, blurry face, multiple girls, crowd, child, loli, petite, text, watermark, signature, long tail

LoRA weight: 1.0. Checkpoint: Nova Furry XL — IL v18.0 A. Steps: 50. Sampler: Euler a. CFG scale: 4.5. Seed: 1802692668. Size: 1024 × 1024. Clip Skip: 2.

Example 3: City skyline with a dark navy jacket

Prompt

shiro_sama, dark navy jacket, white dress, side slit, black thighhighs, black choker, 1girl, solo, city skyline, smile, short tail

Negative prompt

tears, crying, cry, sweat, rain, worst quality, low quality, lowres, jpeg artifacts, bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, extra arms, extra legs, disconnected limbs, multiple tails, detached tail, cat ears, human ears, photorealistic, 3d, cgi, realistic skin, painterly, gradient shading, monochrome, muted colors, underexposed, blurry face, multiple girls, crowd, child, loli, petite, text, watermark, signature, long tail

LoRA weight: 1.0. Checkpoint: Nova Furry XL — IL v18.0 A. Steps: 50. Sampler: Euler a. CFG scale: 7. Seed: 760125420. Size: 1024 × 1024. Clip Skip: 2.

Dataset and caption curation

The dataset contains 598 tagged synthetic images generated with NovelAI 4.5 Curated. Caption curation combined PixAI auto-tagging with manual review, custom tooling, and AI-assisted policy refinement. The tagging dictionary and baseline rules were carried over from my previous Senko-san LoRA projects, then revised and refined for Shiro.

Tag curation used a customized version of toshiaki1729’s dataset editor, modified to support AI-assisted dataset review and tag management. The editor applied the active policy to the raw auto-tagger output and produced curated captions.

AI-assisted review surfaced edge cases, normalized inconsistent terminology, and classified tags as Keep, Reject, or Unknown. The policy was refined and reapplied across the dataset to reduce caption noise and concept bleed.

Training configuration

Training used ostris/ai-toolkit with Illustrious XL v0.1, SDXL architecture, and BF16 precision. The configuration used 2,000 training steps, a batch size of 2, gradient accumulation of 1, and network rank/alpha of 32/32.

The optimizer was Automagic, with a UNet learning rate of 1e-4, a text encoder learning rate of 1e-5, and a cosine learning-rate scheduler.

Tag shuffling was enabled, with the first tag, shiro_sama, preserved. Tag dropout was 0.05, and Clip Skip was 2.