📝 Qwen Image Caption

Qwen Image Caption is the "describe this picture" node from the 24oiduts-ComfyUI pack (Studio42's LCARS Edition). You feed it an image, it runs a real Qwen-VL vision-language model locally, and you get a caption back - the kind of thing people usually reach for BLIP or a cloud API to do. This one keeps everything on your machine.

It's the node to grab when you're building a dataset pipeline: auto-captioning a batch of images before training a LoRA, generating alt-text, or just feeding visual descriptions into a prompt-engineering workflow. If you've ever sat there typing captions by hand for a few hundred images, you know why this exists. The capture is it does this without leaving ComfyUI - your generated images go in, descriptive text comes out, no upload step.

The mechanism

Under the hood it's transformers loading Qwen's vision-language model (the source loads Qwen/Qwen-VL-Chat with trust_remote_code) and building a from_list_format query that pairs your image with the captioning request. Like the other model nodes in this pack, the weights download once and then stay cached in memory, so consecutive calls are fast. The heavy lift is the first run: Qwen-VL is a real multi-billion-parameter model, so expect a multi-gigabyte download and a real GPU to make it comfortable.

What you set

• image - the IMAGE input, straight off any Load Image or VAE decode.
• caption_style - Detailed, Brief, Creative, or Technical. This is the personality switch: Brief for a short tag-style line, Detailed when you want long descriptive prose for dataset training.
• language - English, Chinese, or Auto. Handy if you caption in languages other than English.

That's it - three controls, no hidden state. The output (listed as output in the schema) is the caption text, ready to feed into a STRING input further down the graph.

Where it fits

This is the practical companion to the pack's other vision node, Qwen Vision-Language. The difference: this one is a one-way caption generator with style presets, while Vision-Language takes an arbitrary question. If you want a caption with a particular format, the caption_style dropdown here beats typing a prompt each time. If you want to ask "what color is the sky?", use the other one.

Installing it

Same pack as everything else - 24oiduts-ComfyUI:

• ComfyUI Manager: search 24oiduts, install, restart.
• Manual:

  cd ComfyUI/custom_nodes/
  git clone https://github.com/GeekyGhost/24oiduts-ComfyUI
  pip install -r 24oiduts-ComfyUI/requirements.txt

  then restart ComfyUI.

The transformers dependency is the critical one. Note the GitHub README is an unmaintained Studio42 template that says the project is WIP with no license - treat that as stale docs, the code itself installs cleanly.

Troubleshooting

• Caption comes back as [FALLBACK]... - transformers isn't installed, so the node returned its placeholder instead of erroring. That silent-fallback behavior is a pack-wide quirk; check the console for ⚠️ transformers not available.
• First caption takes ages - it's downloading the model. Let it finish once and it's cached after.
• Out of memory: close other graphs or use a lighter captioner; Qwen-VL is heavy on the VRAM.

If you're already deep in ComfyUI and don't want to ship your training data to a third-party captioning service, this is the easiest local path the pack offers.