Draw Bounding Box Mask

Bounding boxes are great for "where is the thing," but most ComfyUI tools want a mask. DrawBBoxMask is the shape-shifter: it takes a list of BBOX objects and paints them into a MASK on a canvas you size yourself.

The use case that makes it worth knowing: automated inpainting. Detect a face, a watermark, a stray object - get boxes - and DrawBBoxMask turns those boxes into the exact region VAE Encode (for Inpainting) or any inpaint model wants. It's the difference between "here's where the problem is" and "here's a mask that says fix exactly here."

How it works

The node allocates a zeroed float32 tensor of your chosen width × height and fills every box's region with 1.0. Boxes are clipped to the canvas edges, so out-of-bounds coordinates just get trimmed. One canvas, one mask tensor with a batch dimension, white where the boxes were, black everywhere else.

That's the whole mechanism, and it's refreshingly honest - no blur, no feathering, no antialiased edges. It's a hard-edged mask. If your inpaint wants soft transitions, run a blur node after this one (the pack ships a mask-blur-adjacent toolset elsewhere, or any core Comfy blur works on masks).

Inputs and outputs

• bboxes - the list of boxes (the pack's BBOX type) to paint.
• width / height - canvas size in pixels, default 512 each, 64–8192 in steps of 8. Match these to the image you'll inpaint, or the mask won't line up.

Output is a MASK.

Where it fits

Classic chain: detection JSON → ParseBBoxQwenVL (or MaskToBBox from an existing mask) → DrawBBoxMask → inpaint. It also composes multiple boxes into a single mask naturally - several detected regions become one multi-region mask in a single node, which is genuinely useful for "remove every detected object" jobs.

Installing it

It's in Duanyll Nodepack like the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/Duanyll/duanyll_nodepack

or install via ComfyUI Manager and restart. No models, no downloads.

Gotchas

The canvas size is on you. Feed a 512×512 mask into a 1024×1024 inpaint and it silently covers a quarter of the image - nothing will scream at you. Double-check width/height match the target image. Also remember it's a filled box, not an outline: if you only want the region's edge masked, this isn't the node. And because it's hard-edged, expect pixel-perfect boundaries rather than feathered ones - again, blur after if you need it.