📥 Ready-to-Run ComfyUI Workflow Attached! (See
Files->CROWN_MiniMax_H3_Image_Douyin_Chunyu_Yujie_ComfyUI_Workflow.zip)
Includes the pre-configured ComfyUI 0.43+ NativeMiniMax-H3-Image6-Step PDD Workflow JSON + a Drag-and-Drop Workflow PNG (drag straight into ComfyUI, or drag any of the 8 showcase images below directly into ComfyUI!).
⚡ Lossless SVD-Slim Architecture (147.9 MB): Distilled from our 600-stepRank-32master via exact FP64 Reduced-QR + SVD down to a compact147.9 MB(Rank-16 BF16) while preserving98.07%spectral energy across all 208MiniMax-H3-ImageDiT & TokenRefiner layers.
⚠️ 中文必读(含开箱即用 ComfyUI 工作流压缩包):请在上方下载区同时下载 LoRA 本体与CROWN_MiniMax_H3_Image_Douyin_Chunyu_Yujie_ComfyUI_Workflow.zip工作流压缩包(内含 JSON 工作流与支持直接拖入 ComfyUI 的 PNG 样片,下方 8 张展示原图也全部内置了可拖拽 ComfyUI 工作流)!使用官方 6 步ManualSigmasPDD 蒸馏采样表,仅需 6 步(CFG 1.0) 即可直出1664×2496(415 万像素)超清写真!
✨📸 CROWN Douyin Pure-Desire & Alluring Yujie Influencer Photo Realism | MiniMax-H3-Image
[English]
Trained by NineCrowns ([CROWN]) at native 1536×2304 (3.54 MP) on an 8×NVIDIA B200 cluster for MiniMax-H3-Image across 304 human-eye curated, zero-watermark portraits (151 close-up facial beauty portraits + 153 three-quarter/full-body alluring OOTD fashion shots), each individually captioned with Qwen3-VL and encoded through Qwen3-VL-32B (Layer 50) + Ostris De-Distillation Training Adapter (minimax_h3_image_training_adapter_v1.safetensors):
- Top-Tier East Asian Douyin Pure-Desire & Cold-Luxury Yujie Bone Structure (
dy_chunyu_yujie): Delivers sculpted V-line jawlines, high cranial-top hair volume, luminous doe/cat eyes with crystal-clear tear-duct catchlights, delicate aegyo-sal peach blush, glossy lips, and authentic skin pore micro-texture at up to1664×2496(4.15 MP) resolution. - Dual-Bucket Close-Up + Full-Body Alluring OOTD Mastery: Balances 50% close-up beauty headshots and 50% three-quarter/full-body S-curve fashion OOTD shots (lace corsets, satin slip dresses, off-shoulder knitwear, velvet evening gowns, tailored mini skirts) with clean 5-finger hand anatomy.
- 6-Step Native PDD Ultra-Fast Inference: Works natively with
MiniMax-H3-Image's 6-step Piecewise-Distilled Diffusion (PDD) schedule (CFG = 1.0,euler) — no extra turbo LoRA required!
[中文说明]
由 NineCrowns ([CROWN]) 经多智能体集群逐张人工肉眼严审 304 张原生 1536×2304 超清无水印正片(151 张近景怼脸纯欲/御姐神颜特写 + 153 张中远景/全身显身材辣妹穿搭 OOTD),配合 Qwen3-VL 逐图独立摄影级标注与官方去蒸馏适配器(minimax_h3_image_training_adapter_v1),在 8×NVIDIA B200 集群上精炼并经 FP64 正交 QR-SVD 无损瘦身而成的 MiniMax-H3-Image 旗舰写真 LoRA(147.9 MB):
- 顶级抖音纯欲风 & 清冷财阀御姐建模骨相(
dy_chunyu_yujie):精准呈现高颅顶蓬松发量、精致小 V 脸、水光小鹿眼/微醺狐狸眼、清澈泪阜眼神光、纯欲卧蚕腮红与真实微毛孔冷白通透肤质,在1664×2496(415 万像素)下发丝睫毛根根分明。 - 近景怼脸 + 全身 S 曲线穿搭双桶均衡:1:1 兼顾怼脸美妆特写与七分身/全身穿搭 OOTD(蕾丝束腰、真丝吊带裙、一字肩针织、丝绒晚礼服、高腰包臀短裙),腰臀比优越且五指修长自然。
- 原生 6 步 PDD 极速直出(无需外挂加速 LoRA):直接适配
MiniMax-H3-Image官方 6 步 PDD 采样时间表(CFG = 1.0,euler),6 步即可极速出图。
🛠️ ComfyUI Workflow & Recommended Settings / ComfyUI 工作流与参数配置
- Diffusion Model (底模 -
ComfyUI/models/diffusion_models/):minimax_h3_image_pruned_bf16.safetensors(Full BF16) orminimax_h3_image_pruned_int8_convrot.safetensors(16GB–24GB VRAM) - Text Encoder (文本编码器 -
ComfyUI/models/text_encoders/):qwen3vl_32b_minimax_h3_int8_convrot.safetensorsorqwen3vl_32b_minimax_h3_bf16.safetensors(CLIPLoadertype:minimax) - VAE (
ComfyUI/models/vae/):minimax_h3_image_vae_fp16.safetensors - LoRA Weight (推荐权重):
0.85(Recommended range:0.80 ~ 0.95) - Sigma Shift (
MiniMaxH3SigmaShift):shift_video = 1.0, shift_audio = 1.0 - 6-Step
ManualSigmasSchedule (必填 6 步 PDD 时间表):
1.0, 0.9675236940383911, 0.8875485062599182, 0.75, 0.5327739119529724, 0.23200887441635132, 0.0 - Sampler & Guider (采样器):
SamplerCustomAdvanced+BasicGuider(CFG = 1.0) +KSamplerSelect ("euler") - Resolution (推荐分辨率):
1536×2304or1664×2496(2:3 Vertical Portrait) - Trigger Words (触发词):
dy_chunyu_yujie, East Asian top-tier Douyin beauty influencer portrait
Description
[EN] v1.0 1536×2304 Lossless SVD-Slim Rank-16 BF16 release (147.9 MB, retaining 98.07% spectral energy of the 600-step Rank-32 master) + Ready-to-Run ComfyUI 6-Step PDD Workflow ZIP. Recommended weight: 0.80–0.95 (0.85 optimal).
[中文] v1.0 1536×2304 无损 QR-SVD 瘦身正式版(147.9 MB,保留 600 步 Rank-32 母模 98.07% 谱能量),附带开箱即用 ComfyUI 6 步 PDD 官方工作流压缩包(支持拖入图片直接加载工作流)。推荐权重:0.80–0.95(最佳 0.85)。







