Load Qwen Model

If you give PosterCraft a lazy one-line prompt, the results will be lazy. That's what this node exists to fix. LoadQwenModel points the pipeline at a Qwen LLM that rewrites your short idea into a detailed, vivid poster brief before Flux ever sees it - the "intelligent prompt rewriting" step from the upstream PosterCraft paper. A good recap prompt is the difference between "poster with text" and a poster that actually looks designed.

What it actually does

Like the other Load nodes in this pack, it's a pass-through: you type a model_path, it emits that same string under the QWEN type, and the PosterCraft node does the heavy lifting. The rewrite logic lives inside the generator - Qwen gets a carefully worded system prompt telling it to preserve every detail of your prompt, expand it into under 300 words, and handle text elements strictly: use the exact title you provided, never invent new text, describe each text's font, color, and position. If the prompt has no text, it says so explicitly. That discipline is a big part of why PosterCraft's text rendering reads as "designed" rather than "AI gibberish."

The default model_path is Qwen/Qwen3-8B - Alibaba's open-weights 8-billion-parameter model, Apache 2.0 licensed, which is a nice fit for a pack that already rides on FLUX's non-commercial license.

The gotcha that will quietly ruin your posters

Here's the thing nobody warns you about: the generator only loads the Qwen model if os.path.exists(qwen_model_path) is true - meaning the path must point at a local folder that exists on disk. The default Qwen/Qwen3-8B is a HuggingFace repo id, and unless you've downloaded that exact model into a folder with that exact name relative to where ComfyUI runs, the check silently fails and the recap step is skipped.

So with the default settings, your PosterCraft runs on your raw prompt, and the node never tells you. To actually get prompt rewriting you need to:

huggingface-cli download Qwen/Qwen3-8B --local-dir Qwen/Qwen3-8B

run from ComfyUI's root folder - or download it anywhere and type the absolute path into this node. If the model fails to load for any other reason, Qwen quietly falls back to using your original prompt, so a broken recap is invisible except in the console logs.

Practical notes

There's no loading preview here - the model materializes only when the PosterCraft node executes, alongside Flux and the transformer. Qwen3-8B is roughly 16 GB in BF16, so enabling it pushes VRAM well past the ~24 GB Flux alone needs; on a 24 GB card you'll likely be forced to choose between the LLM and the poster. If that's your situation, skip Qwen (leave it at the default, which does nothing anyway) and hand-write richer prompts - the transformer still works without it. Install the pack via Manager (search "PosterCraft") or git clone https://github.com/Yuan-ManX/ComfyUI-PosterCraft, restart, and you'll find this under the PosterCraft category.