中文說明文件 (README_zh.md)

A powerful set of ComfyUI custom nodes designed for batch image processing and advanced Vision-Language Model (VLM) interrogation using the Qwen-VL family (Qwen2-VL, Qwen2.5-VL, Qwen3-VL).

• Batch Image Loading & Renaming: Automatically read all images from a folder, rename them sequentially (e.g., image_1.png, image_2.png), and save them to a specified output directory.
• Multi-Model Support:
  • Qwen-VL: Seamless support for Qwen2-VL, Qwen2.5-VL, and Qwen3-VL models.
  • JoyCaption: Integration with fancyfeast/llama-joycaption for high-quality, natural language captions or Stable Diffusion tags.
• Flexible Workflow Modes:
  • Batch Processing: Process entire directories of images with auto-saving.
  • Single Image: New independent nodes (Single) to process individual images directly from your workflow.
• Advanced Generation Control:
  • Adjustable Max New Tokens, Resolution Control, Temperature, and Seed.
  • JoyCaption specific: Control caption type ('Descriptive', 'SD Prompt'), length, and tone.
• Image Resizing Options:
  • Pixel Resize: Scale based on the longest edge.
  • Megapixel (MP) Resize: Scale to a target total pixel count (e.g., 1.0MP).
  • UI Protection: Smart mutually exclusive toggles in the interface.
• Auto-Saving: Automatically saves the generated captions as .txt files matching the image filenames.

1. Clone this repository into your ComfyUI custom_nodes directory:

   cd ComfyUI/custom_nodes
   git clone https://github.com/tackcrypto1031/tk_comfyui_imageVL.git

2. Install the required dependencies:

   cd tk_comfyui_imageVL
   pip install -r requirements.txt

   Note: This requires transformers, qwen_vl_utils, accelerate, and huggingface_hub.

3. Restart ComfyUI.

This node handles the input images.

• source_path: Directory containing your original images.
• output_path: Directory where renamed images will be saved.
• filename_prefix: Prefix for the renamed files (default: image_).
• resize_px / img_px: Toggle pixel-based resizing (longest edge).
• resize_mp / img_mp: Toggle megapixel-based resizing (total pixels).

This node analyzes the images and generates descriptions.

• model_id: Select the desired Qwen-VL model from the dropdown (Supports Qwen2, Qwen2.5, Qwen3).
  • Models will be automatically downloaded to tk_comfyui_imageVL/models if not present.
• prompt: The instruction for the model (e.g., "Describe this image detailedly.").
• min_pixels / max_pixels: Control the resolution for the vision encoder.
• max_new_tokens: Maximum length of the generated text.
• temperature / seed: Generation parameters.

Process a single image passed directly from another node (IMAGE type).

• image: Input image connection.
• model_id: Select Qwen-VL model.
• prompt: Instruction for the model.
• Returns: STRING (generated text).

Batch processing node for JoyCaption models, designed for natural language captions or SD prompts.

• joycaption_model: Select model (e.g., fancyfeast/llama-joycaption-beta-one-hf-llava).
• caption_type:
  • Descriptive: Formal, natural language description.
  • Stable Diffusion Prompt: Tag-based format with quality boosters.
• caption_length: constrain the output length (very, short - very long).
• user_prompt: Override the internal system prompt with your own instruction.
• cache_model: Keep model loaded (recommended for batch).
• resize_px / img_px: Toggle pixel-based resizing.
• resize_mp / img_mp: Toggle megapixel-based resizing.
• enable_captioning: Toggle to enable/disable caption generation. If disabled, the node acts as a resize-only batch loader.

Single image version of JoyCaption.

• image: Input image connection.
• joycaption_model: Select model.
• caption_type / caption_length: Format controls.
• Returns: STRING (generated text).

This node is kept for workflow compatibility. Text saving is now handled automatically by the Interrogator nodes (TK QwenVL Interrogator / TK JoyCaption Interrogator), which save a .txt file alongside the processed image in the output_path.

1. Load Images: Connect TK Batch Image Loader to TK QwenVL Interrogator.
2. Generate: Connect TK QwenVL Interrogator (texts, filenames) to TK Text Saver.
3. Run: Press "Queue Prompt" to process the entire folder in batch.

Based on the Qwen-VL architecture. Inspired by various ComfyUI community contributions.

This project is licensed under the MIT License - see the LICENSE file for details.