UForm Gen2 Qwen

Small is a feature. UForm-Gen2 Qwen is a 500-million-parameter vision-language model built for exactly two jobs - image captioning and visual question answering - and it's fast enough that you'll wonder why you ever loaded a 7B VLM to describe a cat. This node wraps it, auto-downloads it, and keeps the whole thing as frictionless as a node can be.

What it is

The pack's UForm-Gen2 Qwen node, running the unum-cloud/uform-gen2-qwen-500m model. The README's description is accurate and worth taking literally: "an extremely fast small generative vision-language model primarily designed for Image Captioning and Visual Question Answering." It's not trying to be a reasoning powerhouse - it's trying to be the captioner you actually run, and at that job it punches way above its weight.

The inputs

Two, and that's it:

• image - what you want described.
• question - a multiline STRING. "Describe this image in detail" gets you a caption; "What color is the car?" gets you an answer. The model is a Q&A model, so the phrasing genuinely shapes the output.

Output is a single STRING with the model's response.

How it works

On first use the node automatically downloads the model files into models/LLavacheckpoints/files_for_uform_gen2_qwen - no dropdowns, no GGUF pairing, no clip projector to mismanage. That's a genuinely nicer setup than the pack's llama.cpp nodes: it's the same auto-download pattern as the Kosmos-2 node, and it means the only barrier is disk space and patience on the first run. Subsequent runs use the cached model, so per-query latency is tiny - which is the entire point of a 500M model.

Install

Pack-level install, nothing extra:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI_VLM_nodes

Restart ComfyUI (or Manager → "VLM Nodes"). Python 3.9+. The model downloads itself on first use, so just make sure you have a couple GB free.

Gotchas

First run is a download, not a hang - give it time and watch the console for progress. If it fails partway, delete the partial files_for_uform_gen2_qwen folder and re-run. And set your expectations: it's a captioner, not a prompt engineer. For turning captions into generation prompts, feed its STRING output into CreativeArtPromptGenerator or LLavaPromptGenerator rather than expecting the small model to do both jobs. Use it where speed matters - batch captioning, quick VQA, iterating on a workflow - and reach for the big LLaVa nodes when you need depth.