👁️ Qwen Vision-Language

Qwen Vision-Language is the flexible one in this pack's vision trio. Qwen Image Caption describes an image in a preset style; Qwen Image-to-Text overlays captions onto the image itself. This node lets you ask an arbitrary question about an image - "is this a cat?", "what does the text on the sign say?", "describe the lighting direction" - and get an answer back, powered by a real Qwen-VL vision-language model running entirely on your machine.

It's a question-answering node, not a captioning one. If you're building a quality-check pass over a batch of generations, asking the model "is there a second person in this image?" beats eyeballing a thousand thumbnails. If you want a question-and-answer loop inside a workflow - where the answer feeds a decision node that re-routes your graph - this is the hinge you need.

The mechanism

Like the pack's other AI nodes, it's transformers loading a Qwen-VL model (Qwen-VL-Chat with trust_remote_code in the source), then building the image+question query with the tokenizer's from_list_format helper. The model generates an answer, which comes out alongside an image output. Weights download once and are cached in memory via the pack's shared model cache, so the pain is front-loaded: multi-gigabyte first download, and a real GPU to keep inference pleasant.

The inputs that matter

• image - the picture you're asking about.
• question - the actual query. The default ("What do you see in this image?") is a fine starting point, but the node shines when you ask something specific and answerable.
• model_variant - Qwen-VL or Qwen-VL-Chat. Qwen-VL-Chat is the one to reach for; the plain Qwen-VL is the base model.
• max_tokens - cap on answer length, up to 1024. Long enough for detailed answers.

The output port (listed as output) carries the generated answer text. The pack-wide caveat applies: if transformers isn't installed, this node returns a [FALLBACK] Analysis: This appears to be an image. Question: ... placeholder instead of raising an error. It looks plausible at a glance and it's completely fake - check the console for ⚠️ transformers not available before trusting any answer.

How it differs from the other vision nodes

• Qwen Image Caption - preset styles, no free-form question.
• Qwen Image-to-Text - caption overlaid onto the image as text, good for visual QA or thumbnails.
• This one - true Q&A. The answer is text you can route onward, and the question is anything you want.

Installing it

Same pack, 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 one that must actually be present. Don't be put off by the GitHub README - it's a stale Studio42 template claiming the project is WIP with no license; the code installs and runs normally.

Troubleshooting

• Answers are boilerplate - the fallback placeholder fired; install transformers and restart.
• Slow first question - model download. It's cached afterward.
• Out of memory - Qwen-VL is heavy; close other graphs or switch to a lighter caption path.

For turning images into structured, programmatic answers inside your workflow, this is the node in the pack that actually delivers what it advertises.