CROWN 2D Anime-to-Real 1:1 Photorealistic Redraw | MOSS-Image Edit

🚀 Run Online (No Local GPU Needed) / 云端免显卡一键在线运行：
🌐 RunningHub Global (.ai): 🎬 One-Click AI WebApp | ⚡ Cloud ComfyUI Workflow | LoRA Model
🇨🇳 RunningHub 国内站 (.cn): 🎬 一键 AI 应用 | ⚡ 配套云端工作流 | LoRA 模型页
🤗 Hugging Face Mirror: https://huggingface.co/S3ckn11g0r/CROWN-MOSS-Image-Edit-Anime2Real-1to1

---

📥 Ready-to-Run ComfyUI Workflow Attached! (See Files -> CROWN_MOSS_Image_Edit_Anime2Real_1to1_ComfyUI_Workflow.zip)
⚠️ CRITICAL ComfyUI Setting (Why default template seems to do nothing):
In ComfyUI (>= 0.43.0), you MUST set edit_format = "ai-toolkit" on BOTH the Positive and Negative TextEncodeMOSSImageEdit nodes! (ComfyUI's built-in template defaults to Moss-Image-Edit, which disables Qwen3.5-9B VLM vision tokens).
✅ Works 100% Standalone: Load ONLY CROWN_MOSS_Image_Edit_Anime2Real_1to1_V1_r64.safetensors at strength 0.95 — NO second edit LoRA is needed!
⚠️ 中文必读：请在下载区获取配套 ComfyUI 工作流压缩包！在 ComfyUI 中，正负向两个 TextEncodeMOSSImageEdit 节点的 edit_format 必须同时切换为 ai-toolkit（默认 Moss-Image-Edit 会关闭视觉塔特征导致无效），本 LoRA 权重设 0.95 独立运行即可，无需叠加第二个 Edit LoRA！

🚀 Run Online (No Local GPU Needed) / 云端一键在线运行：
🌐 RunningHub Global (.ai): https://www.runninghub.ai/model/public/2120440207093829634
🇨🇳 RunningHub 国内站 (.cn): https://www.runninghub.cn/model/public/2120453063944265730
🤗 Hugging Face Mirror: https://huggingface.co/S3ckn11g0r/CROWN-MOSS-Image-Edit-Anime2Real-1to1

---

🎨➡️📸 CROWN 2D Anime-to-Real 1:1 Photorealistic Redraw | MOSS-Image Edit

[English]
Trained by NineCrowns ([CROWN]) at native 1024×1536 on an 8×NVIDIA B200 cluster (Rank 64 / Alpha 64) using dual-pathway MOSS-Image Edit conditioning (Qwen3.5-9B VL vision tokens + Flux2-VAE reference latents initialized from moss_image_edit_lora_rank_64_bf16 across 514 curated 2D_ANIME -> 1:1_REAL pairs).
Converts flat 2D Japanese cel-shaded anime illustrations into natural, photorealistic live-action human portraits while preserving 1:1 pose, hairstyle, facial expression, outfit design, and background composition:

1. True Real-Human Anatomy & Skin Micro-Texture (crown_anime2real, A2R): Resculpts stylized 2D anime eyes, nose, and chin into authentic East Asian / human facial bone structure with natural skin pores, realistic eyelashes, and individual hair strands—avoiding plastic 2.5D doll faces.
2. 1:1 Structural, Expression & Outfit Fidelity: Preserves exact lip state (closed-mouth vs open-mouth smile), bangs/hairstyle geometry, fabric folds, accessories, and lighting layout from the 2D reference image.
3. Native ComfyUI Split Q/K/V Format: Exported in standard ComfyUI diffusion_model.layers.*.proj.{q,k,v,o,ffn_up_gate,ffn_down} key layout for plug-and-play loading.

[中文说明]
由 NineCrowns ([CROWN]) 在 8×NVIDIA B200 集群上以原生 1024×1536 分辨率、Rank 64 / Alpha 64 满血规格精炼的 MOSS-Image Edit 二次元 1:1 像素级重绘成真人 LoRA（融合 Qwen3.5-9B 视觉塔 Token 与 Flux2-VAE 参考图潜变量双路控制，基于 514 组严审去重 2D 动漫 -> 1:1 真人 配对精调）：

1. 真实人类骨相重塑与毛孔级肤质（crown_anime2real, A2R）：将平涂二次元大眼与纸片五官自然重塑为真实人类眼眶、鼻梁骨相、皮肤微肌理与根根分明的真实发丝，拒绝 2.5D 塑料假娃娃脸。
2. 1:1 表情、发型、穿搭与构图锁定：严格保持原二次元参考图的嘴部开合状态（抿嘴/微笑/露齿笑）、刘海发辫走向、服饰材质纹样与背景光影构图。
3. 原生兼容 ComfyUI：已拆分为标准 ComfyUI proj.{q,k,v} 独立键位格式，开箱即用。

---

📊 Recommended Settings / 推荐推理参数

• Base Model (底模): moss_image_sft_bf16.safetensors + qwen3.5_9b_bf16.safetensors (VLM vision enabled) + flux2-vae.safetensors
• Resolution & Sampler (分辨率与采样): 1024×1536 (or 1024×1024), 28–30 steps, shift: 5.0, CFG: 3.2–3.5
• LoRA Weight (推荐权重): 0.90 ~ 1.00 (All 8 unseen OOD showcase comparisons generated at 0.95)
• Trigger Words (触发词): crown_anime2real, A2R