Load Bounding Boxes

This is the other half of Save Bounding Boxes: it reads a .cbbox file out of output/bboxes/ and hands you the boxes back as a BOUNDING_BOX, plus an INT frame count. The pitch is simple - detection is the slow, VRAM-hungry step, and once you have the boxes you can delete the detector from your graph entirely and still re-render as many times as you like.

That matters more in video than in stills. A SAM3 track over a few hundred frames is a real wait, and the output is a few kilobytes of numbers. Locking those numbers into a file turns a twenty-minute experiment into a ten-second one.

How it works

The loader opens the JSON, checks the envelope (format: "mvex_bboxes", a bboxes payload), and normalizes it the same way the saver did. If the file holds a single broadcast box, you get that box straight back. If it holds per-frame data, you get the frame list - and this is where the frame controls come in.

skip_first_frames, frame_load_cap and select_every_nth mirror the semantics of the video-loader nodes people already know from VideoHelperSuite: drop the first N frames, take at most N from the start point, then step through with a stride. So skip_first_frames: 10, frame_load_cap: 50, select_every_nth: 2 gives you 25 boxes, matching a video loaded with the same window. That's the point - you window the video and the boxes identically and the frame N box still belongs to frame N.

The frame_count output is the count after slicing, which is what you want to sanity-check against the video's frame count or feed into whatever number field downstream cares. And the loader is honest about cache invalidation: it reports the file's mtime as its changed-value, so if you overwrite clip_01.cbbox with a fresh detection pass, the node re-runs instead of serving you the old boxes from cache.

The inputs you actually set

• bbox_file - dropdown built from output/bboxes/. Shows (none) when the folder is empty, which is the state you'll see on first install.
• subfolder - must match whatever the saver used. Defaults to bboxes.
• skip_first_frames / frame_load_cap / select_every_nth - the window. Defaults of 0, 0, 1 load everything.

Outputs are bboxes (BOUNDING_BOX) and frame_count (INT). Send the boxes into whatever consumed the detector before - a detailer, a crop, a mask builder - and use frame_count when you need to keep the video and the boxes in step.

Install

Manager, search the pack title, or:

cd ComfyUI/custom_nodes
git clone https://github.com/swanjohn99/comfyUIcostumNodes

Restart ComfyUI. No dependencies to install and no models to download; the pack's pyproject.toml declares none, and the loader only uses folder_paths and json.

Where people get burned

The dropdown is built during node definition, which means a file you saved thirty seconds ago in the same session may not be listed yet. The pack's frontend extension papers over this - it fetches the folder list and auto-selects the newest file whose name matches the video you loaded - but if the selection looks stale, refresh the browser and it re-reads the folder. On an older frontend you may have to reload the page after saving.

skip_first_frames is invalid on a broadcast file, and the node says so rather than silently ignoring you: a single box with no per-frame structure cannot be sliced. If you need a window, save per-frame boxes.

Two more. .pt files are rejected deliberately - ComfyUI's model scanner false-flags .pt in output folders as missing models, which is why this pack invented .cbbox - so pointing this at an old pickled box file won't work, and nothing here will convert it for you. And the loader only looks under output/<subfolder>; it will not read an arbitrary path off your disk, so copy the file into output/bboxes/ first if that's where it lives.

If you select (none) and run anyway, you get a plain "run Save Bounding Boxes first" - which is usually the correct diagnosis.