⭐ Star Qwen Regional Prompter

Regional prompting is where diffusion starts to feel like a paint-by-numbers: "sky in the top half, meadow in the bottom, a red car in the lower right." Traditional CLIP can't really do that - a prompt is a prompt, spatial intent gets lost. Star Qwen Regional Prompter is built for the Qwen2-VL CLIP + Qwen-Image stack, and it does the thing cleanly: it splits your image into four quadrants and lets you write a separate prompt for each one, with a background prompt covering everything else.

This is squarely a Qwen workflow node, not a general SDXL helper. It's for when you're editing or generating with Qwen-Image and need composition-level control beyond a single sentence. The node uses a "combined mode" of spatial descriptions plus grounding tokens under the hood - that's the mechanism that makes the model actually place content in the right quadrant instead of just reading your text as vibes.

How it works

You feed it a CLIP (your Qwen text encoder), a background_prompt, and the target image_width / image_height. The four optional inputs - region_upper_left, region_upper_right, region_lower_left, region_lower_right - each take a prompt for their quadrant; leave one empty and it's skipped. The node encodes everything into a single CONDITIONING output that you hand to your sampler's positive input.

The quadrant framing is fixed, which is both the simplicity and the limitation: you get four quarters, not arbitrary regions. For free-form masks or boxes you'd want a fancier regional setup, but for the "compose the frame in quarters" case this is dead simple.

Inputs and outputs

• clip - your Qwen CLIP encoder. Required.
• background_prompt (default A beautiful scene) - covers any area a region doesn't claim.
• image_width / image_height - target resolution; both step by 8 and range 64–8192.
• region_upper_left / upper_right / lower_left / lower_right - the four quadrant prompts. Leave empty to skip a region.

Output: one CONDITIONING socket. Wire it into your sampler's positive input.

Install

Part of the Starnodes2024/ComfyUI_StarNodes pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

Restart and search star. No per-node dependencies - but you do need a Qwen-Image model and its CLIP installed in ComfyUI for the node to be useful at all.

Where people get burned

The classic mistake is treating the quadrants as strict and expecting the model to respect them like a mask - they're guidance, not hard cuts, so content will bleed across quadrant edges. Keep regional prompts descriptive of what belongs there, and let the background prompt carry the scene glue. Also remember this is Qwen-specific: plugging in an SDXL CLIP gets you an immediate error, not regional control. And if you're on a non-Qwen edit pipeline, this node isn't the tool - it's one part of the pack's Qwen conditioning stack, not a universal regional prompter.