ArchAi3D_Mask_To_Position_Guide

Qwen-VL models genuinely struggle with "the second object from the left." Point them at a scene and ask them to modify one of several similar items and they'll pick wrong about half the time. The workaround that this pack commits to is position guides: a reference image where each region you care about is wrapped in a numbered rectangle, so the model can say "region 2" with confidence. ArchAi3D_Mask_To_Position_Guide builds that guide image from a mask, automatically.

How it works

You feed it a mask - white where regions are, black elsewhere (from a segmentation node, a hand-painted mask, whatever). The node detects the connected regions, draws a numbered rectangle around each one, and renders a clean guide image. The clever part is the control you get over how the numbers map to the image, because that determines how you write the prompt:

• numbering_order - left_to_right, right_to_left, top_to_bottom, bottom_to_top. Region 1 is wherever this says first, so if you number left-to-right you must write "region 1 = the leftmost thing" in your prompt.
• rectangle_color / number_color - pick from the presets (red, blue, green, yellow, magenta, cyan, white, black) or custom_hex with a rect_hex_color / num_hex_color field (defaults FF0000).
• line_thickness (default 2), number_size (default 18), padding (default 5) - cosmetics around the boxes.
• background_color - the guide's background; default is custom_hex with 0F0F0F, a dark backdrop that makes the colored boxes pop.

Outputs:

• guide_image - the numbered guide, meant to be fed to the pack's Multi Image Text Encoder V3 alongside the regional text.
• region_count - how many regions were detected (a good sanity check that your mask parsed the way you thought).
• bbox_list - the bounding boxes as a JSON string, if you need the geometry downstream.

The workflow it feeds

Typical chain: segment or paint regions → Mask to Position Guide → build region text with ArchAi3D_Extract_Region_Text ("region 1: 'repaint the wall', region 2: 'swap the flooring'") → encode guide + text together → Qwen edit. The guide gives the model spatial anchors; the numbered text gives it the instructions per anchor. It's the pack's answer to the regional-prompting problem the KB documents - and notably it replaces pixel coordinates with numbered regions, which is exactly the thing Qwen-VL actually responds to.

Install and notes

cd ComfyUI/custom_nodes
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git
cd ComfyUI-ArchAi3d-Qwen && pip install -r requirements.txt

The node uses OpenCV (cv2) for region detection, which is in the Metric3D dependency group - if you installed only the pack's core deps, pip install opencv-python covers it. ComfyUI Manager → search "ArchAi3d Qwen" works too. Keep the numbering order consistent between the guide and your text, and remember: free personal, paid commercial, like everything in the pack.