Draw Mask On Image

Sooner or later you need to see the mask. Not as a separate black-and-white image, but overlaid on the picture: did the detector catch the whole face or just the chin? Did the feather reach the hair? Draw Mask On Image is that, for a whole batch, in one node.

What it actually does

For each frame it blends the image towards color, weighted by m = mask × alpha:

rgb × (1 − m) + color × m

An RGBA frame keeps the larger of its own alpha and m. That's it - a straight linear blend towards a target colour, per frame, no compositing node in between.

The details that make it practical:

• A mask of a different size than the image is scaled to the image with nearest-neighbour. So a mask you made at half resolution still works.
• Fewer masks than frames: they repeat in order. More masks than frames: extras are ignored. Batch mismatches don't blow up, which is the single most common reason mask workflows die.
• A missing or empty mask returns the image unchanged. An empty image batch passes through. You can leave it in the graph while you build the rest.
• Frames are blended one at a time into one preallocated output, and the inputs are never copied - so a 100-frame 1080p batch doesn't triple its own memory footprint doing this.

Inputs and outputs

• image - RGB or RGBA.
• mask - the MASK.
• color - a string widget, default 0, 0, 0. This is the one to read carefully: it takes r, g, b or r, g, b, a comma-separated, where a value above 1 is read as 0–255 and anything else as 0–1. So 255, 0, 0, 128 is a half-opacity red, and 1, 0, 0 is also red, and 1, 1, 1 is white - not 1/255 grey. It also accepts #rgb, #rgba, #rrggbb, #rrggbbaa. Alpha is the opacity, and it's 1 when you leave it out.
• device - cpu (default) or gpu.
• Output: images, the blended batch, same shape as the input.

The alpha-in-color thing is the nicest part of the design. 255, 0, 255, 90 is a loud magenta wash you can still see the image through - much better for judging a mask than a solid fill or a 50/50 blend you didn't choose.

What it's for

Debugging, mostly. A magenta overlay tells you in one glance what a SAM or YOLO detection actually grabbed, and the same overlay on a video batch tells you whether a detector is jittering frame to frame - flicker is invisible in a still mask and obvious when you scrub an overlay.

Beyond debugging, it's a competent compositor for flat-colour work: paint a region solid for a graphic, lay a colour grade over the subject only, or produce a matte-region preview for someone who doesn't want to read a black-and-white PNG.

If what you want is an actual composite with real image content, this isn't it - use ImageCompositeMasked or run the region through a second sampler. This node paints a colour, and it's honest about that.

Install

The usual for this pack: ComfyUI Manager, search ComfyUI-BCNodes, install. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/0xBeycan/ComfyUI-BCNodes
pip install -r ComfyUI-BCNodes/requirements.txt

Restart and search BCNodes in the node library (it sits under BCNodes/mask). requirements.txt is opencv-python, postfx and caption-audit; this node needs none of them to load. If ComfyUI still has an old ComfyUI-AutoBypass folder from the pack's earlier name, delete it - the two will clash.

Gotchas

The color parser is strict and will raise rather than guess. A colour in a form it doesn't recognise, or one whose value count doesn't match the image's channels, stops the run with a message about what to fix. That's better than silently painting black, but it means rgb(255,0,0) - not one of the documented forms - is an error, not a colour. Stick to the documented formats; #rrggbb is the least error-prone.

Second: because a too-short mask list repeats in order and a too-long one gets truncated, a shape mismatch fails quietly. If your overlay is covering the wrong thing on frame 30, check the mask batch length before you check the mask itself.

Third: device matters more than you'd expect on video batches. Blending 300 frames at 4K on CPU is measurable; on GPU it's free. The default is conservative on purpose - flip it yourself.