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

Files with SHA256: c7008ff98dc66a78ec85e63c89402e8d86f427169f42185458aad3c07badb8ea