ArchAi3D_Position_Guide_Prompt_Builder

Getting Qwen-Image-Edit to put multiple objects in a scene with a specific layout is where prompt writing breaks down - "put the boy on the left, the man next to him" is a coin flip. The trick this node uses is old-school and brutally effective: you hand the model a reference image with numbered red rectangles drawn where things should go, and the prompt tells it "object 1 goes in rectangle 1, object 2 in rectangle 2." The model reads the image, maps your descriptions to the boxes, adds the objects, and the prompt also tells it to remove the red rectangles and numbers when it's done.

What it does

ArchAi3D_Position_Guide_Prompt_Builder turns a list of object descriptions into that placement prompt. You feed it:

• object_descriptions - slash-separated descriptions, each mapped to a rectangle in order. "boy seats on sofa / man seats on sofa / woman seats on sofa" means rectangle 1 gets the boy, rectangle 2 the man, rectangle 3 the woman.
• prompt_template - the cleanup strategy. standard (combined removal command) is the default; standard_explicit splits the removal into separate commands if rectangles or numbers leak into the output; strong_removal adds an IMPORTANT prefix for stubborn cases; no_removal keeps the boxes visible for debugging; minimal is just the mappings; custom uses your own template.
• system_prompt_template - seven options, from user_validated (the recommended one, a "your working prompt" with a 3-stage cleanup) through three_stage (explicit Read → Add → Clean process) to none. The preserve_everything and minimal_change options are there for when the scene matters more than the placement.
• custom_template / custom_system_prompt - escape hatches that use {MAPPINGS} and {ADDITIONAL} placeholders when you pick the custom presets.
• additional_instructions - free-form extras like "man is close to boy" or "warm lighting, photorealistic style".
• removal_at_start and removal_at_end - where the removal command goes, and how it's worded (combined, rectangles_only, numbers_only, both_separate). The defaults put a combined removal at the end.

Outputs

Two strings: formatted_prompt - the complete prompt, ready for your Qwen encoder - and system_prompt, which is empty if you selected none.

How it works under the hood

The node is a structured template composer. It builds the mapping list from your slash-separated input, wraps it in the chosen template, and appends the removal instructions at start and/or end. The real work happens in the model: Qwen-Edit 2509's vision side reads the numbered rectangles from the guide image, and the natural-language instruction maps your descriptions onto them. It's the same "the image is part of the prompt" trick that makes Qwen Edit good at instruction editing in general - you're just making the instruction machine-readable by giving it coordinates to hang on.

When to reach for it

Composition shots where layout matters and you're tired of rolling the dice: seating charts, product-lineup shots, "three characters in a scene" work. It pairs with an image-editing workflow where you first draw the guide (numbered rectangles on a copy of the image), then feed the original plus the guide into the encoder alongside this prompt.

Install + notes

Pack install - ComfyUI Manager ("ArchAi3d Qwen") or:

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

Restart, find it under ArchAi3d/Utils. The biggest gotcha is leaky cleanup: if red boxes or digits show up in the output, step the template up - standard → standard_explicit → strong_removal. That escalation ladder is the entire troubleshooting guide, and it works.