Kuromaruart Art Style is a high-quality style LoRA trained on the distinctive artwork of Kuromaru (@kuromaruart), a popular R-18 / NSFW digital artist known for sensual, highly detailed anime-style illustrations with a strong emphasis on character appeal, dynamic lighting, and erotic themes.
Base Model
Trained on Illustrious XL 2.0 (illustriousXL20_v20.safetensors) for excellent compatibility with SDXL workflows.
Key Training Details
LoRA Settings: Rank 32, Alpha 16 with dropout (network_dropout 0.1, rank/module dropout 0.05) for better generalization.
Dataset: Curated images from Kuromaruart's signature style.
Resolution: Bucketed around 1280x1280 with enable_bucket and no_upscale for high-detail training.
Optimizer & Scheduler: Adafactor with cosine_with_restarts (3 cycles), warm-up, and SNR-based noise/huber loss for clean, high-quality convergence.
Other Optimizations: FP16 mixed precision, xFormers, gradient checkpointing, cache_latents, clip_skip=2, etc.
Recommended Usage
Trigger words: kuromaruart style, by kuromaruart, or simply kuromaruart
Strength: 0.8 – 1.2 (start at 1.0 and adjust; works well up to 1.4 for stronger stylization)
Best with: Illustrious XL or other strong anime/SDXL models. Pairs excellently with character LoRAs.
Negative prompt suggestions: low quality, blurry, deformed, bad anatomy (standard SDXL negatives work great)
Style Characteristics
This LoRA faithfully reproduces Kuromaruart's signature aesthetic:
Lush, sensual rendering with smooth skin, soft yet defined shading, and glossy highlights.
Beautiful lighting — dramatic rim lights, volumetric effects, and appealing color grading.
Expressive, attractive faces with large expressive eyes and seductive expressions.
Highly detailed hair, clothing textures, and fabric folds with excellent drape and translucency.
Strong in NSFW / erotic scenes, tentacles, monster girls, restrained poses, and atmospheric lewd compositions, while also performing well on SFW portraits and character illustrations.
The model was trained for 15 epochs (~4206 steps) with caption dropout and shuffling to maintain flexibility without overfitting.