QwenVL Prompt Enhancer

You type "a cat on a beach," and this node hands back a paragraph with lighting, composition, camera, and mood spelled out - the kind of dense prompt modern models actually respond to. It's a text-only node: no image input, no vision. Just a local Qwen3 language model rewriting your prompt into something richer before it hits your sampler. Despite living in a "QwenVL" pack, this one never looks at a picture.

Here's the detail that makes it click: the LLMs doing the enhancing are the same family that encodes prompts for today's models. Z-Image runs on a Qwen3-4B text encoder. Flux 2 Klein uses Qwen3. Anima uses Qwen3-0.6B. So when this node expands your prompt with a Qwen3 model, you're essentially pre-writing in the encoder's native dialect. That's a nicer story than piping your prompt through some unrelated chatbot API - it's local, keyless, and shaped by the same model lineage that's going to read the result.

How it works

It's a plain instruction-following LLM with a system prompt bolted on. You give it your rough prompt and pick a style - the style selects a hidden system instruction ("expand this," "make it cinematic," "add technical camera specs") and the model rewrites accordingly. No magic, just a well-prompted small language model running on your GPU instead of behind a paywall.

The inputs that matter

• prompt_text - your input prompt, the thing to be enhanced. Multiline.
• enhancement_style - six presets: Enhance, Refine, Creative Rewrite, Detailed Visual, Artistic Style, Technical Specs. "Enhance" is the sensible default (flesh it out); "Refine" is lighter-touch; "Creative Rewrite" takes the most liberty. Match the style to how far you want it to stray from your original.
• model_name - 21 options. The pure text models are the ones to start with: Qwen3-0.6B (tiny and fast, fine for basic expansion), Qwen3-4B-Instruct-2507 (the stronger general pick), and qwen3-4b-Z-Image-Engineer, a community fine-tune aimed specifically at writing Z-Image prompts. The Qwen-VL models are also listed and run here in text-only mode.

If the presets don't fit, custom_system_prompt lets you write your own instruction and take full control of the rewrite. The sampling trio - temperature (default 0.7), top_p, repetition_penalty - controls how wild the rewrite gets; turn temperature down if it's inventing details you don't want. The single output is ENHANCED_OUTPUT (a STRING); wire it straight into your CLIP/text encoder in place of the raw prompt. keep_model_loaded and seed behave as everywhere else in the pack.

How to install it

Same pack as the rest. ComfyUI Manager: search ComfyUI-QwenVL, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/1038lab/ComfyUI-QwenVL.git
cd ComfyUI-QwenVL
pip install -r requirements.txt

Restart ComfyUI. The text models download from HuggingFace on first use - Qwen3-0.6B is small, but the 4B models are a few GB, so the first run pauses to fetch them.

Common issues

The honest caveat: this is an LLM, so it will confidently invent. Ask for "more detail" and it may add a red scarf, golden-hour light, and a Leica that were never in your idea - great when you want inspiration, annoying when you had a specific image in mind. Lower the temperature and lean on "Refine" over "Creative Rewrite" if it's drifting, and always read the output before trusting it. Longer isn't automatically better; a bloated prompt can dilute the tokens that mattered. Beyond that, the usual pack notes apply: a 401 on first download is a stale HuggingFace token (clear it, these are public models), and if you're short on VRAM, the 0.6B model runs almost anywhere while the 4B and larger picks need real headroom.