Shiro-sama_NAI-V4.5-Curated (シロ様)

High-fidelity character LoRA for Shiro-sama (シロ様) from The Helpful Fox Senko-san, trained on FLUX.2 Klein 9B Base.

This LoRA was built from a high-quality synthetic dataset of 598 captioned images generated with NovelAI 4.5 Curated. It is designed for txt2img character generation, img2img, inpainting, and character replacements, with an emphasis on character likeness, clean linework, outfit flexibility, and descriptive natural-language prompting.

Trigger word

shiro_sama

Include shiro_sama in your positive prompt, followed by a description of the outfit, activity, setting, and visual style you want.

What this model does best

• Recognizable identity: Shiro’s reddish-pink eyes, long white hair, fox ears, short fluffy white tail, pink cheek markings, and catlike smile.

• Outfit flexibility: Her canonical white-dress ensemble, jackets, casual clothing, seasonal outfits, and other wardrobe choices.

• Descriptive prompting: Detailed control over clothing, expressions, poses, framing, lighting, and environments.

• Image editing: Direct instructions for character changes or targeted corrections to an existing image.

Recommended generation settings

• Base model: FLUX.2 Klein 9B Base

• Trigger word: shiro_sama

• LoRA strength: Start at 1.0; try 0.7–1.0 to balance likeness and flexibility

• Sampler: Euler (euler)

• Scheduler: Simple (simple)

• CFG scale: Start at 7 for txt2img and img2img, then adjust for the task

• Steps — faster generation: 28

• Steps — balanced: 40

• Steps — higher-detail option: 50

• Resolution: Start at 1024 × 1024

• Negative prompt: Completely empty

Do not use a negative prompt with this LoRA. Leave the negative-prompt field completely empty.

The examples below use 35 steps, with CFG 4.5 for txt2img and CFG 10 for targeted img2img edits. These are the actual example settings, separate from the general starting recommendation above.

Prompting guidance

Use direct, specific instructions. Start with shiro_sama, then describe the image in clear sentences.

For txt2img, specify her clothing, expression, pose, camera angle, setting, and lighting. Reinforce important character features when needed, such as her white hair, white fox ears, reddish-pink eyes, pink cheek markings, and short fluffy white tail.

For img2img and inpainting, clearly describe what should change and what should remain unchanged. A focused instruction can be enough for a small correction; broader character replacements benefit from a fuller description of her appearance.

When color accuracy matters, describe Shiro’s hair, ears, and tail as white, with lighting-dependent highlights and shadows. Her eyes should remain reddish-pink, and her cheek markings pink.

Use short tail to reinforce the intended tail length. Avoid detailed eyes if it causes her black pupils to disappear.

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 garment descriptions; shiro_sama remains the character trigger.

For alternate outfits, explicitly describe the replacement clothing. Retain descriptions only for the default pieces you want to keep. When adding a jacket or cardigan, omit conflicting garment descriptions and specify the new clothing directly.

Generation examples

All four examples use FLUX.2 Klein 9B Base, a LoRA weight of 1.0, 35 steps, Euler (euler), Simple (simple), and 1024 × 1024 resolution.

Examples 1 and 2 have empty negative prompts. I forgot to clear the negative prompt for Examples 3 and 4, and those results did not come out as nicely. They are included for reference, but do not copy that part of their settings: leave your negative prompt empty.

Example 1: Autumn walk — txt2img

Prompt

shiro_sama shiro_sama with her fluffy white tail, white fox ears, reddish-pink eyes, pink cheek markings, red diamond pendant, and catlike mouth, taking a cheerful autumn walk, looking directly at the viewer with bright, relaxed eyes and a small, confident smile. Frame her closely from above her ears to mid-thigh, with her face as the main focus. Her body angles along the sidewalk while her head turns toward the viewer, caught mid-step with one shoulder slightly forward. She wears a short moss-green cardigan ending exactly at her waistband, hanging open over a cream mock-neck top, and a high-waisted plum pleated skirt. Simple knit texture, softly rounded sleeves, and a small amber leaf brooch. She holds a takeaway coffee below her chest, clear of her face, while her free arm swings naturally. Her hair and skirt move gently with her stride. Behind her on image-right is a softly lit Halloween shop window bordered by small orange and purple lights, with a few ghost and bat stickers. Inside, pumpkins and a plush ghost sit on a purple carpeted display. An autumn tree beside the shop frames the upper-left with orange-gold leaves. Warm late-afternoon sunlight catches her hair and shoulder, with gentle light across her face. Keep the shop softly detailed and its warm interior dimmer than her face, preserving a bright outdoor golden-hour atmosphere. Anime, anime coloring.

Workflow: txt2img. LoRA weight: 1.0. Steps: 35. Sampler: Euler. Scheduler: Simple. CFG scale: 4.5. Seed: 308583758. Size: 1024 × 1024. Negative prompt: Empty.

Example 2: Targeted hand correction — img2img

Prompt

shiro_sama shiro_sama's held out hand on the right side of the image should only have 4 fingers and a thumb

This example uses a short, focused editing instruction to request a correction to the hand on image-right.

Workflow: img2img edit, with one input image. LoRA weight: 1.0. Steps: 35. Sampler: Euler. Scheduler: Simple. CFG scale: 10. Seed: 2076514806. Size: 1024 × 1024. Negative prompt: Empty.

Example 3: Autumn walk — txt2img with an accidental negative prompt

This generation uses the same positive prompt as Example 1, with a different seed and the accidental negative prompt reproduced below.

Workflow: txt2img. LoRA weight: 1.0. Steps: 35. Sampler: Euler. Scheduler: Simple. CFG scale: 4.5. Seed: 326839520. Size: 1024 × 1024.

Negative prompt: The shared accidental negative prompt below. Leave this empty when generating your own images.

Example 4: Targeted hand correction — img2img with an accidental negative prompt

Prompt

shiro_sama shiro_sama's held out hand on the right side of the image should only have 5 fingers.

Workflow: img2img edit, with one input image. LoRA weight: 1.0. Steps: 35. Sampler: Euler. Scheduler: Simple. CFG scale: 10. Seed: 1063669291. Size: 1024 × 1024.

Negative prompt: The shared accidental negative prompt below. Leave this empty when generating your own images.

Accidental negative prompt used in Examples 3 and 4

The following is preserved only to document the settings used for those two examples. It is not a recommended 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

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.