CV Deep Features

CV Deep Features runs an ONNX backbone over your frames and gives you one embedding row per frame. That's it. No training, no gradients, no new model weights - a frozen network used as a ruler.

Why would you want that in ComfyUI? Two real reasons. First, image similarity without a model download: embed a folder of references, embed a generated image, sort by cosine distance, and you have "find me the ones that look like this" without a CLIP or a DINOv2 in sight. Second, training a small classifier on your own frames - the pack's CV Train Classifier takes these embeddings and fits a cv2.ml model, which is the classical-CV route to "is this a defect / a face / my cat".

The honest framing: cv2.dnn cannot backpropagate. Nothing here is a LoRA, nothing here touches your diffusion model, and for most style- or content-matching jobs a CLIP vision embed does more. Reach for this when you want small, deterministic and CPU-friendly, or when you want to see how the classical pipeline fits together.

How it works, and the one setting that matters

The frames go through blobFromImage (resize to a square, per-channel mean subtraction, scale, BGR→RGB) and are pushed through the net. Crucially, you pick which layer to tap. layer_name should be an interior bottleneck - for SqueezeNet 1.1 that's squeezenet0_concat7, which yields 512-dim embeddings. Leave it blank and you get the model's final output, i.e. class logits, which transfer considerably worse than an interior activation. Then the activation map is pooled into a single row: average over space per channel (position-invariant, the standard embedding) or flatten the whole thing if you want layout retained.

The node pins the net to OpenCV's classic DNN engine, because the newer fused graph engine mis-handles interior taps. Classic engine layer names carry an onnx_node! prefix, which the node adds for you when the bare name isn't found.

Inputs and outputs

Required: image (a whole IMAGE batch gives one row per frame; a plain NPARRAY counts as one frame), model (an .onnx from ComfyUI/models/onnx), layer_name, input_size (224 for the ImageNet backbones), scale (default 1/255), swap_rb (leave on - OpenCV frames are BGR, the backbones want RGB), and pooling.

Optional, and only worth touching when your backbone isn't ImageNet-trained: mean_r/mean_g/mean_b (default the ImageNet means, 0 disables) and std_r/std_g/std_b (default the ImageNet stds, 1 disables).

Outputs: features - the (N, D) float32 embedding matrix, which wires straight into CV Train Classifier or CV Stack Feature Classes - plus dims (D), count (frames embedded) and success.

Install

Part of comfyui_cv (bmad4ever/comfyui_cv) - search "ComfyUI CV" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_cv
pip install "opencv-contrib-python-headless~=5.0.0.93"

Restart afterwards. Python ≥ 3.12 and a V3-API ComfyUI. Models are not bundled - download the ONNX backbone yourself and drop it in ComfyUI/models/onnx; the pack's model_sources.txt records URLs and licences for the ones its workflows use.

Where people get burned

• success = false and an empty (0, 0) array. A wrong layer name or an empty batch does that, deliberately, instead of raising. Gate anything downstream with if/else on success. Also note this node can only return the batch it was given - grey or BGR NPARRAYs are treated as one frame.
• It raises on a build with no classic engine. OpenCV 5.1 removed that engine, and tapping an interior layer becomes impossible - the call fails loudly rather than quietly handing you logits. The pinned 5.0.0.93 wheel is the supported configuration; that's not laziness, it's the one the pack is tested against.
• Installing a plain opencv-python wheel. All four OpenCV distributions share one site-packages/cv2, so a non-contrib wheel overwrites a contrib one and the contrib nodes disappear. Fix with python tools/repair_opencv_contrib.py --check then --apply. This is the single most common way a working CV install breaks in the wild - the "cv2 DLL load failed / nodes won't load" threads are almost all this class of conflict.

Practical tip: keep the batch small. This runs through cv2.dnn on the CPU, so 200 frames of 224x224 is fine and 200 frames of 1024 is your evening gone.