SAM3 Draw BBox

SAM3 (Meta's Segment Anything 3) is promptable, and its favorite prompt is a box. Which means half of building a SAM3 workflow is just getting the boxes right - from a VLM, from a YOLO pass, from the interactive editor, from whatever. Sam3DrawBBox is the "show your work" node: it takes an image and a BBOX collection and paints a red rectangle around every box, so you can actually see what you're about to segment before you burn GPU time on it.

In: images (IMAGE), bboxs (BBOX), stroke_width (INT, 1–64, default 4). Out: images - the same tensor with boxes drawn on top. There's nothing clever beyond that, and there doesn't need to be. It's a debug/visualization tool, and it's genuinely useful when your boxes come from an auto-detector you don't fully trust.

Why the format tolerance matters

The BBOX type is shared across this ecosystem (kijai's KJNodes and the easy-sam3 pack both speak it), and the ecosystem is sloppy about exactly what it hands you. This node tries to accept all of it:

• KJ-format dicts - {startX, startY, endX, endY}, what the interactive editor in this pack emits.
• {x, y, w, h} dicts - treated as top-left plus size.
• Raw lists - a 4-element list is read as xyxy unless x2 <= x1, in which case the node guesses it's xywh.
• Normalized coordinates - if everything's ≤ 1.0, it scales up to the image size.
• JSON strings and per-batch lists - a string gets parsed, and a list with one entry per image gets applied image-by-image; otherwise one set of boxes applies to the whole batch.

That's a lot of guessing baked in, and it mostly works - which is the honest description: mostly. When boxes don't draw, it's usually because your source used a format the heuristics didn't recognize, so the node silently skipped it rather than crash.

The one knob

• stroke_width (INT, 1–64, default 4) - the only parameter you'll touch. 4 is a good middle ground; on high-res images you may want a heavier line, on previews a thinner one. All boxes render in red; there's no color option, so if you need per-class colors this isn't your node.

Installing it

The pack-level install:

• ComfyUI Manager → search comfyui-easy-sam3-tools → Install → restart.
• Or:

cd ComfyUI/custom_nodes
git clone https://github.com/SuLU-K/comfyui-easy-sam3-tools

Then restart ComfyUI. No requirements.txt, no model downloads - drawing rectangles needs only Pillow, which ComfyUI already has. The node appears under Test Nodes/SAM3, the pack's honest label for its personal-utility origins.

Troubleshooting

• Nothing draws. Either bboxs is empty/None, or your format fell through the heuristics. Check the shape of what you're feeding: if it's a list of four numbers, is it xyxy or xywh? If it's a dict, does it have the keys above? A degenerate box (zero width/height) is also dropped silently.
• Boxes in the wrong spot. The node clamps everything to image bounds, and normalized values get multiplied up by image size. If your boxes came in already-pixel-scaled and under 1.0 in magnitude, they'll be treated as normalized - check your upstream's coordinate space.
• Every image in a batch gets the same boxes unless your input is explicitly per-batch. That's the default, and it's usually what you want for a preview.

If your goal is just "confirm the detector found the right thing," this is the node. If you need to author boxes, you want Sam3InteractiveBBoxEditor in the same pack instead.