## 📖 Overview / 概览
Enhance MOSS image prompt optimization with QWEN2.1 PE. Two ComfyUI workflows included:
- T2I — refine a rough text prompt into a detailed generation-ready caption
- I2I/WD14 — input an image, extract WD14 tags via TaggerPlus (PixAi-v1.0), then rewrite an optimized prompt
通过QWEN2.1 PE增强 MOSS 提示词优化，包含两个工作流：
- T2I — 将粗略文本提示词优化为可直接生成的详细描述
- I2I/WD14 — 输入图像，通过 TaggerPlus（PixAi-v1.0）提取 WD14 标签，再重写优化提示词
Files: moss_image_prompt Optimization_T2I_V0.1.json & moss_image_prompt Optimization_I2I_WD14_V0.1.json
---
## ✨ Features / 功能
- QWEN2.1 PE enhanced prompt rewriting / QWEN2.1 PE 增强提示词重写
- T2I + I2I/WD14 dual modes / T2I 与 I2I/WD14 双模式
- Auto-download PixAi-v1.0 tagger model on first run / 首次运行自动下载 PixAi-v1.0
- Tested on RTX 3060 12G, ~3 min per optimization / RTX 3060 12G 测试，单次约 3 分钟
---
## 📦 Requirements / 要求
- ComfyUI-TaggerPlus (I2I/WD14)
- ComfyUI_Comfyroll_CustomNodes (CR Text series)
- ComfyUI-Easy-Use (easy seed / clearCache)
- rgthree-comfy (Lora Loader Stack)
- comfyui-custom-scripts (ShowText)
- ResolutionSelector
MOSS Image models (from HuggingFace: Comfy-Org/MOSS-Image): / MOSS Image 模型（来自 HuggingFace: Comfy-Org/MOSS-Image）：
- qwen3.5_9b_uncensored_int8_convrot.safetensors → models/clip/ (T2I)
- qwen3.5_9b_qwen_image_2.1_pe_i2i_heretic.int8_convrot.safetensors → models/clip/ (I2I/WD14)
- moss_image_sft_int8_convrot.safetensors → models/unet/
- flux2-vae.safetensors → models/vae/
---
## 🚀 Quick Start / 快速上手
1. Install ComfyUI, Manager, and all custom nodes above.
2. Download MOSS Image models into correct folders.
3. Load the .json workflow in ComfyUI (Load → select file).
4. Queue Prompt. (I2I/WD14: PixAi-v1.0 downloads automatically on first run.)
1. 安装 ComfyUI、Manager 以及上述所有自定义节点。
2. 将 MOSS Image 模型下载到相应的文件夹中。
3. 在 ComfyUI 中加载 .json 工作流（“加载”→“选择文件”）。
4. 提交提示词。（I2I/WD14：PixAi-v1.0 会在首次运行时自动下载。）
---
## ⚠️ Notes / 注意事项
- First-run download: I2I/WD14 auto-downloads pixai-tagger-v1.0 . Ensure stable internet.
首次运行： I2I/WD14 会自动下载 pixai-tagger-v1.0，请确保网络稳定。
- Filename match: Update model names in CLIPLoaderUNETLoaderVAELoader if your files differ.
文件名匹配： 若模型文件名不同，请在加载节点中更新。