ComfyUI_QwenVL_PromptCaption

Leverages Qwen 3.5/3/2.5 VL for prompt inversion & caption generation in ComfyUI

---

重要说明 | Important Note

❌ 插件不自动下载模型，可复用 ComfyOrg 提供的 qwen_2.5_vl_7b.safetensors，也可手动下载其它Qwen VL模型。
❌ This plugin does not auto-download models. It can reuse qwen_2.5_vl_7b.safetensors provided by ComfyOrg, or manually download other Qwen VL models.

---

节点 | Nodes

1. Qwen XX VL Caption: image/video prompt inversion
   Qwen XX VL Caption：图片/视频提示词反推
2. Qwen XX VL Batch Caption: Batch image prompt inversion (folder input)
   Qwen XX VL Batch Caption：目录批量图片提示词反推
3. Ovis 2.5 Run: Run Ovis 2.5 model
   Ovis 2.5 Run：运行 Ovis 2.5 模型
4. ASID_Caption: Run ASID Captioner model
   ASID_Caption：运行 ASID Captioner 模型

---

安装方法 | Installation

a. Via ComfyUI Manager
通过 ComfyUI Manager 安装
b. Manual install:
手动安装：

1. Copy the plugin folder to ComfyUI/custom_nodes/
   复制插件目录至 ComfyUI/custom_nodes/
2. Update dependency: transformers>=4.57.0 (>=5.2.0 for Qwen3.5)
   更新依赖：transformers>=4.57.0（Qwen3.5需>=5.2.0）

---

使用方法 | Usage

1. Download the model
   下载模型
2. Edit prompt templates (optional)
   编辑指令提示词（可选）
3. Adjust node inputs
   调整节点输入参数
4. Click "Run"
   点击运行

---

模型说明 | Model Notes

• 模型读取路径：ComfyUI 的 text_encoders 目录（需手动放置已下载模型）。
  Model path: ComfyUI's text_encoders folder (place downloaded models manually).

复用 ComfyOrg 模型 | Reuse ComfyOrg Model

To reuse qwen_2.5_vl_7b.safetensors:
复用 qwen_2.5_vl_7b.safetensors 步骤：

1. Create a FOLDER in ComfyUI/models/text_encoders
   在ComfyUI/models/text_encoders中创建一个文件夹
2. Rename the model file to model.safetensors and move it into the FOLDER
   将模型文件重命名为 model.safetensors并移入创建的文件夹
3. Add required config files (from Qwen 2.5 VL's official Hugging Face repo)
   添加必要配置文件（取自 Qwen 2.5 VL 官方 Hugging Face 仓库） https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct

✅ No extra disk usage – model remains usable for ComfyUI's Qwen Image/Edit model.
✅ 无额外硬盘消耗，不影响原模型用于 ComfyUI 的 Qwen Image/Edit模型。

直接下载官方模型 | Direct Download

Download Qwen 2.5/3 VL official repo from Hugging Face, then place it in text_encoders.
从 Hugging Face 下载 Qwen 2.5/3 VL 官方仓库，直接放入 text_encoders 目录即可。

https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct

https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct

https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct

国内也可从网盘下载：https://pan.quark.cn/s/b3975e789c3c

Support Ovis 2.5 models
新支持 Ovis 2.5 模型

https://huggingface.co/AIDC-AI/Ovis2.5-2B

https://huggingface.co/AIDC-AI/Ovis2.5-9B

Support ASID Captioner models
新支持 ASID Captioner 模型

https://huggingface.co/AudioVisual-Caption/ASID-Captioner-3B

https://huggingface.co/AudioVisual-Caption/ASID-Captioner-7B

---

自定义提示词 | Custom Prompts

Now you can input instruction directly, or
现在可以直接输入指令，或者
Edit prompts.txt in the custom_nodes folder (follow the existing format):
修改插件目录下的 prompts.txt 文件（参考原有格式）：

• Support multiple prompts
  支持多条提示词
• The nodes will use the last prompt matching the language
  自动读取对应语言的最后一条提示词

---

模型精度建议 | VRAM & Precision Recommendations

显存 (VRAM)	推荐精度 (Recommended Precision)
6-8GB	Qwen 2.5 VL 7B (4bit) / Qwen 3 VL 8B (4bit) / Qwen 3 VL 4B (8bit)
10-16GB	Qwen 2.5 VL 7B (8bit) / Qwen 3 VL 8B (8bit) / Qwen 3 VL 4B (bf16)
16GB+	bf16 (full precision)

---

参数说明 | Parameter Notes

max_side

• Pre-scales the image's longer side to this size
  预缩放图片长边尺寸
• Larger values may reduce processing speed
  设置过大会导致速度下降

keep_model_loaded

• Use True to Keep model in VRAM for consecutive prompt inversion tasks
  连续进行提示词反推时选 True
• False won't impact performance during batch node run
  批量节点选False仅在全部图片处理完成后清理模型，不影响过程性能

unload_other_models

• Attempt to unload all models via ComfyUI model management before loading, to avoid VRAM-related loading failures.
  加载新模型前，尝试通过ComfyUI的model management卸载所有模型，以避免因剩余显存不足导致的加载失败。

save_path

• will use image_path to save output if save_path not set
  save_path为空时会使用image_path保存输出