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 Anima Base v1.0 (recommended checkpoint: Anima Turbo v1.1).

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

Trigger word

shiro_sama

What this model does best

Character identity: Maintains Shiro’s recognizable facial features, fox ears, short fox tail, and anime appearance across different outfits and scenes.

Outfit flexibility: Supports her canonical white-dress ensemble as well as jackets, casual clothing, and other alternate outfits.

Descriptive prompting: Supports natural-language descriptions and mixed tag-and-prose prompts for controlling clothing, expressions, poses, composition, lighting, and environments.

Recommended generation settings

Use Anima Base v1.0 or a compatible Anima checkpoint. Both examples below were generated with Anima — turbo-v1.1, at a LoRA weight of 1.0.

For turbo-v1.1, the demonstrated configuration is Euler Ancestral (euler_ancestral / Euler a), Simple scheduler, 15 steps, CFG 1.5, and 1024 × 1024 resolution.

Start with a LoRA weight of 1.0. Try 0.90 or 0.85 when you want more flexibility for outfit changes or scene control.

For Anima Base or another compatible checkpoint, follow that checkpoint’s recommended step count and CFG scale. The 15-step, CFG 1.5 settings shown here apply to the Turbo examples.

Anima does not use Clip Skip.

Prompting

Anima uses Qwen3-0.6B for text conditioning and supports tags, natural-language descriptions, and combinations of both.

Keep shiro_sama as the character trigger, then describe the image directly. For more precise control, specify what she is wearing, where she is positioned, how she faces the viewer, what her hands are doing, and how the scene is lit.

The examples below combine a short set of character and clothing tags with natural-language instructions for the camera position, pose, expression, and environment.

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 descriptions 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. Both examples below use this approach.

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

Avoid detailed eyes if it causes her black pupils to disappear.

Negative prompting

For a minimal negative prompt, start with:

worst quality, low quality, blurry, deformed, text, watermark

Add targeted negatives when needed for ears, tails, anatomy, clothing, or other recurring problems. The examples below include the complete negative prompts used for those generations.

Example 1: Shared park bench beneath a street lamp

Prompt

masterpiece, shiro_sama, dark blue jacket, white dress with a side slit and yellow moon logo, black thigh-highs, black choker, and short tail. Use a first-person view from someone sitting together with shiro_sama on the same park bench at night beneath a street lamp. Include the viewer’s knees and one hand resting on their thigh in the lower foreground, with shiro_sama seated immediately beside them, close enough that their shoulders nearly touch. Angle the camera sideways toward her from the neighboring seat, with the bench slats continuing beneath both people. Her body faces out toward the park while she turns her head toward the viewer with a slight tilt, meeting their gaze. She holds a light blue popsicle with its tip between her lips, paused mid-bite, her eyes widened and eyebrows raised in surprise. Warm street-lamp light illuminates her face and the shared bench against the cool, dark park.

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

Workflow: txt2img. LoRA weight: 1.0. Checkpoint: Anima — turbo-v1.1. Steps: 15. Sampler: Euler Ancestral (euler_ancestral). Scheduler: Simple. CFG scale: 1.5. Seed: 1257711859. Size: 1024 × 1024.

Example 2: Torii gates above the nighttime city

Prompt

shiro_sama, dark blue jacket, white dress with a side slit, black thigh-highs, black choker, and short tail. Place her alone in the center of the frame, walking a few steps ahead of the viewer through a series of torii gates high above the city at night. Her body faces forward along the path while she turns her head back over her shoulder, looking directly at the viewer with a confident, playful, catlike smile as she leads them onward. One foot rests on an ethereal white, translucent walkway while the other steps forward. The walkway stretches through the clouds, with successive torii gates receding into the distance. Far below, the illuminated city is visible through the translucent path and gaps in the clouds.

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

Workflow: txt2img. LoRA weight: 1.0. Checkpoint: Anima — turbo-v1.1. Steps: 15. Sampler: Euler Ancestral (euler_ancestral). Scheduler: Simple. CFG scale: 1.5. Seed: 1663584593. Size: 1024 × 1024.

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 Anima Base v1.0, a standard LoRA, and BF16 precision. The character trigger was shiro_sama.

The configuration used 2,000 training steps, a batch size of 2, and gradient accumulation of 1. Diffusion model training and gradient checkpointing were enabled, with checkpoints saved every 200 steps.

The network used a rank/dimension of 32 and an alpha of 32, giving an effective alpha/rank scaling of 1.0.

The optimizer was Automagic, with a diffusion model learning rate of 0.0001 and a cosine learning-rate scheduler.

The dataset used 598 images with .txt captions and a base training resolution of 1024 × 1024. Caption shuffling was enabled, with Preserve Tag set to 1 to retain the first tag, shiro_sama. Caption dropout was 0.05, and latent caching was enabled.