ArchAi3D_Qwen_Watermark_Removal

Set expectations early: this node never touches pixels. It's a prompt builder for Qwen Image Edit - you pick what to remove and where, it hands you a ready-made instruction, and the actual erasing happens in the Qwen edit model further down your graph. The reason that's worth a node at all is that Qwen edit models are ridiculously literal about phrasing. "Remove the watermark" works; "clean up this image" gets you a guess. This node removes the guesswork by templating the prompt from known-good research.

How it works

Two dropdowns, both meaningful:

• text_type - what to remove: watermark, all_text, english_text, chinese_text, logo, brand_mark, ui_elements.
• location - where: anywhere (let the model auto-detect), or bottom_right, top_left, center, and the rest of the corners/edges for precision.

The node maps those into a clean prompt like "Remove the watermark from the bottom right corner of the image" and emits it as its prompt output. There's also a debug_mode toggle that just prints the generated prompt to your console so you can see what you're actually sending.

The interesting output is system_prompt. That's a packaged system instruction the pack's encoder nodes accept directly on their system_prompt input - a full "you are an image cleanup specialist, inpaint the removed area seamlessly, preserve everything else exactly" persona. The tooltip calls it new in v4.0 and it genuinely matters: Qwen-Edit's default system prompt isn't tuned for surgical removal, and the difference between "blend it seamlessly" and "make it disappear" is usually the system prompt.

How you wire it

Image → (Image Scale) → encoder. Text goes to the encoder's text/conditioning input, system_prompt to the encoder's system prompt input, then sample and decode. It slots into the same Qwen-Edit-2509 pipeline the rest of the pack uses - the node's outputs are explicitly tagged "for Qwen Edit 2509."

Inputs that matter

• text_type - start with watermark; switch to ui_elements for screenshot cleanup and logo for brand marks.
• location - anywhere is a surprisingly good default; Qwen auto-detects well. Use a specific corner when the model keeps "fixing" the wrong part of the frame.
• debug_mode - turn it on once, see the exact prompt, understand what you're dealing with.

Outputs: prompt (STRING) and system_prompt (STRING). Both feed into an encoder node.

Install and caveats

Part of the ArchAi3D Qwen pack, so:

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

Or ComfyUI Manager → search "ArchAi3d Qwen" → restart.

Two honest caveats. First, Qwen-Edit re-emits the whole frame, so pixels you didn't ask to touch can shift - it's the model's structural weakness, not this node's, and the KB's qwen-image-edit panel says it first. Second, remember it's a text generator: if you're getting nothing back, check the connection into the encoder, not this node. And the pack's license is free personal, paid commercial - presumably the same for scrubbing stock photos you don't own.