TK JoyCaption Interrogator

If you're captioning a training set in natural language, JoyCaption is the community default - an 8B VLM built from scratch for exactly this job, free, open, and uncensored by design. TK JoyCaption Interrogator is the batch wrapper that points it at a folder, writes a .txt caption for every image, and gives you back the texts. It's the pack's answer for "I want rich prose captions, not tags, and I don't want to ask an API."

How it works

Give it a source_path and an output_path. It loads the JoyCaption model (default fancyfeast/llama-joycaption-beta-one-hf-llava, with the old pre-alpha as the second option), pulls it into tk_comfyui_imageVL/models on first use, then for each image:

1. optionally resizes (same megapixel / longest-edge toggles as the batch loader),
2. builds the caption prompt from your controls,
3. runs inference and writes {prefix}{n}.txt into output_path,
4. copies the image over with its new {prefix}{n} name.

Outputs are texts (LIST) and filenames (LIST), position-matched. Saving is automatic - there's no separate saver step.

The controls that actually shape the captions

• caption_type - Descriptive (formal prose) or Stable Diffusion Prompt (tag-style with quality boosters like "masterpiece, best quality"). This swaps the internal instruction the model follows.
• caption_length - any, very short, short, medium-length, long, very long. It's appended to the prompt, so treat it as a soft constraint, not a hard token limit.
• user_prompt - overrides the built-in instruction entirely. Leave it empty to use the preset for your caption_type.
• max_new_tokens / temperature / top_p / top_k - generation knobs; 512 default is fine for captions.
• cache_model - keep it on. Off unloads the model after every run, and reloading an 8B VLM each time turns a batch into a waiting game.
• enable_captioning - flip this off and the node becomes a resize-and-rename batch loader that writes no captions. It's the escape hatch if you just want the image pipeline.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/tackcrypto1031/tk_comfyui_imageVL.git
cd tk_comfyui_imageVL
pip install -r requirements.txt

Restart ComfyUI after, or use ComfyUI Manager and search tk_comfyui_imageVL. The node leans on transformers>=4.45.0, accelerate, and huggingface-hub from the pack's requirements.

Where people get burned

• It's an 8B model. Beta One is Llama 3.1 8B plus a SigLIP2 vision encoder. It fits a consumer card but it's not instant - a folder of hundreds of images takes a while, and the first run downloads the weights.
• Multi-subject attribution. The author says it himself: when several people are in a frame, VLMs mix up who's wearing what. Auto-caption the bulk, audit the small sets by hand.
• Not the tool for anime tag datasets. JoyCaption's tag modes approximate booru vocabulary, but if you're captioning for Illustrious/Pony/NoobAI, a WD14 tagger emits the exact tags those models were trained on. Use this node when the base wants prose (Flux, Z-Image, most LLM-encoded models) - and its uncensored stance is a real advantage there, since the API captioners refuse exactly the content you might need captioned.

Batch JoyCaption in a node, no API key, no censorship dance, .txt files done for you. That's the pitch, and it mostly delivers.