✨ CROWN Niji Semi-Realistic Aesthetic & Porcelain Skin Boost | MiniMax H3 自训半写实神颜与皮肤质感加强 LoRA
[English]
This is our self-trained Niji Semi-Realistic Aesthetic, Facial Bone Structure & Subsurface Scattering (SSS) Porcelain Skin Enhancement LoRA (Rank 16, Alpha 16, Video-Only) for MiniMax H3 (Ref2VA / I2VA / T2VA).
Trained on a curated 312-image high-resolution Niji semi-realistic portrait & close-up aesthetic dataset (1280x1280, 1024x1792, 1792x1024) with trigger words stripped from captions so the delicate Niji semi-realistic 3D CGI facial bone structure, luminous glass/porcelain skin micro-texture, subtle lower-eyelid blush makeup, and cinematic studio rim lighting are baked directly into the default weights (no trigger word required — active immediately upon loading). Because it is trained with video_only: True and sound: silent still image isolation, it never interferes with MiniMax H3's native stereo audio generation.
[中文介绍]
这是我们针对 MiniMax H3 (Ref2VA / I2VA / T2VA) 专门训练的 Niji 半写实神颜骨相与次表面散射(SSS)冷白瓷肌质感增强 LoRA(Rank 16 / Alpha 16,Video-Only 视频层隔离训练)。
训练时对 312 张高分辨率 Niji 半写实神颜特写训练集进行了画风触发词剥离(Triggerless Captioning),将 Niji 半写实精致面部骨相、微醺眼妆卧蚕、果冻水润唇釉质感与影棚柔焦次表面散射肌肤纹理 100% 焊入 LoRA 默认权重中(无需填写触发词,挂载即刻生效),同时严格隔离音频层,100% 不干扰 H3 原生立体声配乐与人声对口型。
📊 Real Production Test Record / 真实生产测试记录
- Base Model (底模):
minimax_h3_ref2va_bf16.safetensors(37.5 GB Full BF16) +qwen3vl_32b_minimax_h3_bf16.safetensors+minimax_h3_video_vae_fp32.safetensors+minimax_h3_audio_vae_fp32.safetensors - Sampling Steps & Schedule (采样步数与调度):
20 ~ 24 Steps,eulersampler +ManualSigmas,MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0) - LoRA Stack & Exact Weights Used in Our Close-Up Showcase (本片实测挂载的 LoRA 组合与精确权重):
- Self-Trained Niji Aesthetic LoRA (本自训 Niji 半写实神颜质感加强 LoRA):
CROWN_H3_Niji_Semireal_Aesthetic_V2_r16.safetensors— Weight (权重):1.00 ~ 1.10(Recommended range:0.85 ~ 1.20) - 2K Detail Refiner LoRA (2K 毛孔肌理超清精修 LoRA):
minimax_h3_open_regenerate_2k_refiner_rank_16_bf16.safetensors— Weight (权重):0.35 - Step Acceleration LoRA (可选步数加速 LoRA):
minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors— Weight (权重):0.45
- Self-Trained Niji Aesthetic LoRA (本自训 Niji 半写实神颜质感加强 LoRA):
- Trigger Word (触发词): None required (无需触发词,挂载即生效)
Description
v2.0 Video-Only Rank 16 release (6 Epochs / 1,872 Steps on 312 curated Niji semi-realistic portraits). Triggerless activation — enhances semi-realistic facial bone structure, luminous porcelain skin micro-texture, and cinematic lighting while keeping native stereo audio untouched.