Qwen Generation 🐼

If the pack's other node (Qwen Conversation) is the chatty one, this is the one you'll actually wire into a generation workflow. Drop a rough idea in, get a polished natural-language prompt out, feed that straight into your CLIP Text Encode. That's the whole job: one-shot text generation with a local Qwen3 model, nothing sent anywhere, no API key in sight.

And it's a real workflow, not a gimmick. LLM-assisted prompting went from a browser-tab trick to a ComfyUI staple - and it fits the architecture especially well on LLM-encoded models like Z-Image, Flux 2 Klein, and Anima, where the "text encoder" is itself reading your prompt as an instruction. Having an LLM write that instruction is just translating between two things that speak the same language. This node is the Qwen3 take on that, from Alibaba's Qwen family.

How it works

Same engine as the conversation node: a shared processor loads the model with a transformers text-generation pipeline, optionally quantized through BitsAndBytesConfig, and picks CUDA, MPS, or CPU by what it detects on your machine. The difference is the loop - single user turn, no history kept, so the output is just the assistant's reply. The node strips markdown to plain text before returning it, which is a small mercy when your reply comes back full of **bold** and ### headers.

First run downloads the model automatically (default Qwen3-4B, ~8GB fp16 or ~2.5GB at 4-bit) from whichever of Hugging Face or ModelScope is faster, with fallback and retries. It lands in ComfyUI/models/Qwen/Qwen/<model>. Subsequent runs are instant unless you've told it to unload.

The inputs that matter

• model_name - 26 Qwen3 options, 0.6B up to the 235B MoE. Default Qwen3-4B is right for prompt rewriting; go smaller for pure speed.
• quantization - 4-bit (default), 8-bit, or None. For prompt polish, 4-bit is plenty. Pre-quantized FP8/AWQ/GGUF model variants ignore this setting.
• enable_thinking - lets the model reason out loud before answering. Good for gnarly prompt briefs, wasteful for "make this more cinematic."
• max_new_tokens - default 2048, up to 6144. Rewrites rarely need much.

unload_after_generation defaults to on: VRAM is freed after every run, but you pay a reload cost next time. If you're batching, flip it off and let the model sit.

The output

One text string - the clean assistant answer, plain text, markdown stripped. Wire it into a CLIP Text Encode (or any text input). That's the whole point of this node versus its sibling, which returns the full conversation transcript.

Install

Same pack as Qwen Conversation, so one install covers both:

cd ComfyUI/custom_nodes
git clone https://github.com/SXQBW/ComfyUI-Qwen
cd ComfyUI-Qwen
pip install -r requirements.txt

Or search "ComfyUI-Qwen" in ComfyUI Manager. Note the README's stale ComfyUI-Qwen3.git URLs - the repo is ComfyUI-Qwen now. Dependencies are heavy (torch>=2.6, transformers>=4.51, accelerate, modelscope) but most ComfyUI setups already have them. One real trap: bitsandbytes isn't in requirements.txt, yet 4-bit/8-bit quantization needs it - pip install bitsandbytes if loading fails on that.

Common issues

• First run looks hung. It's downloading gigabytes and loading an LLM into RAM. Wait; it prints bilingual progress to the console.
• "bitsandbytes" errors on 4-bit/8-bit - install it (CUDA only; Mac/CPU fall back to fp16).
• Thinking mode eats your token budget - reasoning chains count against max_new_tokens. Disable it for fast rewrites.
• Tag-based models want tags. If you're feeding an anime/Danbooru-style model, explicitly tell Qwen "output comma-separated tags only" - sentence prose is what it defaults to.
• The default prompt is an ad. The built-in example prompt is a Flux rewrite template that name-drops the author's panda persona. Swap in your own brief - that's the real skill here anyway.

Last thing: this runs real Python and downloads from the HF hub, like every LLM node. Give the repo a glance before you install; the ecosystem has had at least one genuinely malicious LLM node, and install is still a trust decision.