CV Array To Numbers

ComfyUI has a second kind of data flow that surprises people the first time they meet it: a socket can carry a list, and everything downstream then executes once per value. That's how Inspire foreach loops iterate over detections, and it's how Basic Data Handling's whole sort/filter/min/max surface works. CV Array To Numbers is the door from array-land into that world: it takes the values in a numpy array and emits them as a ComfyUI list of FLOATs.

Why you'd reach for it

• Per-detection work. A table of N detections feeds a list of N floats, and a downstream chain runs N times - crop the Nth region, process it, save it. That's the loop construct ComfyUI gives you.
• Sort, filter, aggregate. Hand the list to Basic Data Handling's convert node and you get the whole toolkit: max, min, sum, sort, and the ability to pick one and feed it back as a scalar.
• A score column as a gate. Column 4 of a detection table is usually the score. Emit it as numbers, threshold it in the helper pack, and you've built confidence filtering without a dedicated node for it.

The inputs that matter

• nparray - the array to read. NPARRAY only.
• column - for a 2-D array, which column to emit. -1 means every value, row by row, which is the default. Set it to a specific column and you get just that column - the mechanism above.
• max_items (optional) - a safety cap, default 1024, and it's the most important number in the node.

Outputs: value - one FLOAT per value, in order, marked as a list, so the socket fans out as an iterable. count - how many were actually emitted, after the cap.

The thing to understand before you use it

Every value is a separate execution of everything downstream. The pack's own tooltip says it plainly: a whole image here would run the rest of the graph a million times. That's not a performance warning, it's a design contract - and it's why max_items exists, why its default is a modest 1024, and why you should treat this node as "emit a few numbers and loop over them" rather than "dump an array into the graph".

Rule of thumb: this node belongs on a detection table with tens of rows. If you're thinking in thousands, you want vectorised array operations instead - the pack's wrappers do arithmetic across whole arrays in one execution, which is a couple of orders of magnitude cheaper than looping.

Two mismatches that will bite

• INT and FLOAT sockets don't auto-coerce. This node emits floats by definition. A consumer that wants an integer index needs Basic Data Handling's to INT conversion in between, and it will not warn you beyond refusing to connect.
• Rows are not values. If you wanted to iterate over whole rows as arrays - "each detection as a 4-element array" - this is the wrong node. CV Unstack Batch does that job, and the difference matters as soon as the downstream consumer wants more than one number per item.

Install

ComfyUI Manager → ComfyUI CV, or:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_cv

restart, done. Dependencies are opencv-contrib-python-headless~=5.0.0.93, numpy and torch, with Python 3.12+ and a ComfyUI on the V3 node API. Note that the interesting half of this node's life happens in other packs - the README lists ComfyUI-Inspire-Pack, ComfyUI-Custom-Scripts and Basic Data Handling as dependencies of the pack's own example workflows, and those are exactly the consumers you'd wire this into. Install them via Manager if a downloaded workflow opens red.

Traps

• max_items silently truncates. The count output tells you the truth; if it equals your cap, you lost data. Raise it deliberately, knowing what per-item executions cost you.
• column = -1 on a big 2-D array is a flood. Row-major order, every value, one execution each. On anything image-sized, that's the mistake the cap was designed to prevent - don't defeat it.
• Order is row-major. The Nth value is the Nth cell, so a list index only maps back to a detection if your table is in detection order.
• A list output changes the identity of the node it feeds. Downstream nodes that expected a scalar will now run N times and their outputs accumulate. If you see a graph suddenly producing an unexpected batch, look upstream for the node that turned a value into a list - it's usually this one.