☁️BizyAir Point-Box Guided SAM

Meta's Segment Anything Model made "click on the thing you want masked" a real workflow primitive instead of a research demo, and this is BizyAir's version of it. You load an image, click a point on it (or draw a box around a region), and SAM returns a precise mask for whatever you selected - no text prompt, no training, no fiddling with thresholds. It's the most direct, manual way to get a targeted mask, which makes it the right tool when you know exactly which object you want and just need ComfyUI to trace its outline.

This is deliberately the simple entry point in BizyAir's segmentation lineup - the pack shipped it back in October 2024 as "box_point_guided_segment-anything," alongside a separate text-guided variant. If you'd rather describe the object in words instead of clicking on it, that's the other node; this one is for when pointing is faster than typing.

How it works

SAM was built to take exactly these three kinds of prompts - a point click, a bounding box, or text - and turn any of them into a segmentation mask. This node commits to the first two and lets you flip between them with a single toggle. Click a point inside the object you want (a face, a piece of clothing, a hand) and SAM expands outward to find its boundary; draw a box instead and SAM fills in the mask within it. Point clicks are faster for a single clear object; a box is more reliable when the object is oddly shaped or you need to avoid grabbing something overlapping it.

Inputs and outputs

• image - this is an enum populated from files in your ComfyUI input folder rather than a live IMAGE socket, which is the standard convention for nodes that expect you to click directly on the canvas: pick an uploaded image, then interact with it in the node's preview.
• is_point - a boolean, true by default, that switches the interaction mode. true means you're clicking a point; false switches to drawing a bounding box.

It gives you back three outputs: processed_image (the image with the selection visualized, useful for sanity-checking what you actually selected), mask (the actual segmentation mask - this is the one you wire into inpainting, compositing, or a Detailer-style node), and original_image (the untouched source, handy for passing downstream without re-loading it).

Installing BizyAir

Via ComfyUI Manager: search "BizyAir," click install. Or clone it directly:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

then restart ComfyUI. comfy node install bizyair works too if you're on Comfy-CLI. Don't skip the API key step - click the "click to login" link on first use and register. SAM here runs on SiliconFlow's cloud rather than your local GPU, which is actually the point: full-size SAM checkpoints aren't small, and BizyAir's pitch is you don't need them sitting on your disk.

Common issues

A blank or wrong mask almost always traces back to the click/box you gave it rather than the model - SAM is genuinely good at finding object boundaries once it knows roughly where to look, but a point dropped near an edge or a box that clips the object will hand it an ambiguous prompt. If your masks look off, re-check your selection before assuming the node is broken.

If the node fails outright with no useful output, check your API key first - that's the most common failure mode across this entire pack, since almost nothing runs locally. And if you installed BizyAir as a side effect of a downloaded workflow rather than choosing it yourself, that's a known and fairly common experience with this pack (ComfyUI Manager sometimes resolves missing-node errors to a BizyAir equivalent) - it's not a sign anything's wrong with your setup.