Trained on high-quality CGs from the adult visual novel Nie no Hakoniwa.

All showcase images were generated text-to-image (t2i) using this LoRA on the official Anima base v1.0.

• Resolution Note: Works consistently up to 1536x1024. Pushing it higher might occasionally generate black borders, though this is rare.

V1.3

• Adjust the dataset to include more wider shots.

• I tag the male protagonist (Yasaka Kurou) appearing in dataset, its bleeding into male characters should be lessen. May still happen since male characters tend to have less data presence in base model.

• Reduce the total steps for training, should be more flexible to use with other styles.

Recommended Settings

For the most authentic game feel, I highly recommend using the official Anima base v1.0.

• LoRA Weight: 1.0

• CFG Scale: 4.5 - 5.0

• Sampler: ER-sde (best for matching the original CG style) or Euler a (for a smoother, more modern look)

• Steps: 30 - 50 (Higher resolutions require more steps to bring out fine details)

Known Issues (WIP)

• Male Character Bleed: The LoRA currently tends to affect the identity of male characters, making them look similar to the game's protagonist, Yasaka Kurou.

• Female characters are largely unaffected and generate well.

I will try to fix the male character issue in a future update.

Training Specs

• Trainer Tool: Anima-Standalone-Trainer

• Base Model: Anima base v1.0

• Dataset: 162 179 lightly cropped CG screenshots (1920x1080). No regularization images used.

• Captioning: Natural language captions via Gemini 3.5 Flash & Qwen3.6 27B-heretic. Danbooru tagging via wd-vit-tagger-v3 (with manual pruning).

• Resolution: Mixed (512 / 1024 / 1536)

• Buckets: Min 512 | Max 1536 | Steps 64

• Batch Size: 4, 4, 2

• Network: Dim 32 | Alpha 32

• Optimizer: AdamW (Weight Decay: 0.01)

• Learning Rate: 0.00002 (Text Encoder LR: 0 — UNet only)

• Scheduler: Cosine (Warmup Steps: 100)