LayerUtility: QWenImage2Prompt(Advance)

Quick thing to clear up before anything else: despite the name, this isn't Alibaba's Qwen-VL. The README is upfront about it - QWenImage2Prompt is a repackage of ComfyUI_VLM_nodes's "UForm-Gen2 Qwen Node," which wraps unum-cloud/uform-gen2-qwen-500m. That's a 500-million-parameter vision-language model that happens to use a Qwen tokenizer/decoder component internally, not the full Qwen-VL family you'd get from Alibaba directly. If you came here expecting Qwen2-VL-scale reasoning about your image, you'll want one of the newer VLM options instead - this node is small, fast, and good for a quick caption, not a detailed multi-turn description.

The mechanism is dead simple: feed it an image and a question, it runs the UForm-Gen2 model and returns a text answer. No system prompt, no sampling knobs, no seed control - this is the minimal end of the captioning nodes in this pack, which is either exactly what you want (fast, no config needed) or a limitation depending on your use case.

Inputs and outputs, all of them:

• image - your source picture.
• question - defaults to "describe this image," but you can ask something more specific ("what color is the subject's clothing," for instance) and get a direct answer instead of a generic caption.
• text (output) - the model's response, ready to wire into a prompt field or a text-processing node downstream.

That's the entire interface. If you need finer control over generation - temperature, max tokens, sampling - this isn't the node for it; look at LlamaVision or the Gemini-backed nodes elsewhere in the pack instead.

Installing it: ComfyUI Manager, search "ComfyUI Layer Style Advance," or manually git clone https://github.com/chflame163/ComfyUI_LayerStyle_Advance into custom_nodes, then run install_requirements.bat (portable Windows) or pip install -r requirements.txt in your venv, followed by repair_dependency.bat to clean up the common dependency snags. Restart ComfyUI afterward.

The model itself is a real, separate download - it doesn't come bundled with the node pack. Grab it from huggingface.co/unum-cloud/uform-gen2-qwen-500m (or the BaiduNetdisk mirror) and put it in ComfyUI/models/LLavacheckpoints/files_for_uform_gen2_qwen. That exact folder path matters - the node won't find the model if it's sitting one level up or under a differently-named subfolder.

Where people get stuck: most import or load failures on this node trace back to the same place as the rest of the pack - a missing or mismatched dependency from the large requirements.txt this repo ships (it's split off from the original LayerStyle pack specifically because these nodes need heavier dependencies than the base pack). Run repair_dependency.bat first if the node fails to import cleanly. Beyond that, since this is a smaller, older VLM by current standards, don't expect it to handle nuanced or compositional questions well - it's built for straightforward "what's in this image" captioning, and pushing it past that will just get you vague or wrong answers rather than a clean failure.