CV Class Scores Decode

What this is for

Every classification network ends the same way: a vector with one number per class. Useful - but only if you can turn index 207 into a word. CV DNN Pick Output gets you that vector out of a model's blob; CV Class Scores Decode does the last hundred metres: softmax (or not), sort, top-k, attach the class names.

It's a DATA node. It draws nothing. That's deliberate throughout this pack - the data nodes never render, so you choose what becomes a picture. Wire labels_out into Preview as Text, or use top_label as a STRING downstream.

How it works

The score vector arrives as (1, C) or (C,) and is flattened. Then the activation question: auto (the default) applies softmax only when the values lie outside [0, 1], because that's the signature of raw logits. If the model already ends in softmax or sigmoid, the values are probabilities and re-softmaxing them would flatten a confident prediction into mush. softmax and sigmoid force the operation; none never applies it. If your top-k proportions look weirdly uniform, this is the widget you got wrong.

After that it's a descending argsort and a slice - top-k, capped at the vector length so asking for 10 out of 5 classes returns 5 rather than erroring.

Inputs and outputs

Required: scores (the vector from DNN Pick Output) and top_k (default 5).

Optional:

• activation - auto / softmax / sigmoid / none, as above. Advanced widget; the default is usually right.
• labels - class names, one per line, line N = class id N, pasted from CV Load Labels or anything else. Blank means classes are labelled by their numeric id, which is still readable output.

Four outputs:

• labels_out - (K,) string array, best first. Feed Preview as Text.
• scores_top - (K,) float32 scores after the chosen activation, best first. This is the one you'd chart: run it through CV Chart Series and you get labelled bars instead of a ranked list of numbers.
• class_ids - (K,) int32 ids, if you want to key off the index rather than the name.
• top_label - plain STRING of the top-1 name, for wiring into filenames, notes or prompts.

Install

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

Or ComfyUI CV in ComfyUI Manager (publisher bmad4ever). Needs Python ≥ 3.12 and a V3-node-API ComfyUI; opencv-contrib-python-headless~=5.0.0.93 comes with the pack. No model needed for this node - but the model that produced scores has to be an .onnx in ComfyUI/models/onnx, exported for cv2.dnn.

Common issues

• Every score comes back around the same value. You're softmaxing probabilities (or sigmoiding logits). Flip activation to none or auto and look again.
• Labels don't line up. labels is line-indexed from zero, and the file has to match the model's class ordering - a different ImageNet-sorted list gives you confidently wrong names. CV Load Labels reads the same text format the detectors use, so one label file can serve the whole graph.
• The model doesn't load at all. This is the pack's honest weak spot, and its README says so: everything goes through cv2.dnn on principle, the pinned OpenCV's DNN support is limited to older architectures, and a perfectly valid ONNX export can still be unloadable. CUDA is not available in these wheels either - the DNN backend/target selectors exist but are effectively inert on current builds. For anything ComfyUI already does natively, use the native node.
• Contrib nodes vanished. Non-contrib OpenCV wheel clobbered the shared site-packages/cv2; tools/repair_opencv_contrib.py --check / --apply.