Dataset & Training Overview

• Training Settings: 10 Epochs (Total 5,000 steps)

• Target Subjects: 10 main characters from the game (+ 4 child variants)

• Total Images: 100 images (Trained at 512x512 resolution / 262,144 pixels)

Dataset Breakdown:

• 24 images: 4-split facial expression graphics (cropped close-ups)

• 26 images: Character sprites framed in a cowboy_shot

• 33 images: In-game event CGs (illustrations)

• 17 images: AI-generated supplementary assets (via Gemini):

  • 14 character design sheets (Three-view/Orthographic diagrams)

  • 2 clothing variant references

  • 1 extreme close-up of the pin badge and embroidered emblem patch

Known Issues & Limitations

⚠️ CRITICAL DISCLAIMER (Environment & Compatibility)

All testing and behavior verification for this LoRA were conducted exclusively using "Anima Turbo LoRA v0.2". The generation settings used were: 8steps/er_sde/simple/832x1248. Please note that this LoRA has NOT been optimized or tested for non-Turbo, high-step environments. Your results may vary significantly if you use standard sampling settings.

• Element Bleeding (Tattoo & Embroidery): Ririko's sun tattoo and denim embroidery tend to bleed onto other female characters (1girl). If these elements unexpectedly appear on other characters, they can be mitigated using negative prompts.

• Low-Detail Accessories: Fine details on small accessories such as pin badges, emblem patches, and wristbands may appear somewhat blurry or unrefined.

• Hozuki's Face Instability: The Hozuki (法月のとっつぁん) suffers from highly unstable facial consistency.

• Persistent Emblem Patches: The embroidered patch consistently forces itself onto school uniforms (and sweaters), regardless of whether you prompt the patch tag or not. This can be suppressed using negative prompts if necessary.

• Character Part-Borrowing (No Original Sprites): Characters who lacked official face graphics or standing sprites in the original game (e.g., morita_kenichi, nagumo_eri) are often rendered by borrowing facial and clothing features from other similar characters. (e.g., morita_kenichi might mutate into isono_ichirouta, or vice versa).

• Multi-Character Complexity Limits: Although the dataset includes event CGs featuring 4 or 5 characters, due to Anima's model limitations, accurate reproduction of scenes with 4 or more people is not guaranteed. Expect anatomy breakdowns (corrupted fingers, extra limbs, or distorted arms) in crowded compositions.