ComfyUI Node

CV Deep Features

A frozen ONNX backbone as a fixed feature extractor: each frame is forwarded to layer_name (an interior bottleneck, NOT the classifier logits) and the activation map is global-average-pooled into one embedding row. This is transfer learning the OpenCV way - cv2.dnn cannot backpropagate, so the backbone stays frozen and a cv2.ml head ('CV Train Classifier') is trained on the embeddings instead. E.g. SqueezeNet 1.1: layer squeezenet0_concat7 -> 512-dim embeddings. The net is pinned to OpenCV's CLASSIC engine where the build still has one (the fused graph engine mis-handles interior taps); classic layer names carry an 'onnx_node!' prefix which is added automatically when the bare name is not found. A wrong layer name or an empty batch yields success=false and an empty (0, 0) array, never an error - but a build with NO classic engine (OpenCV 5.1 removed it) cannot tap an interior layer at all, and that RAISES rather than quietly returning the classifier logits.

By bmad4ever·Created 3 months ago·Updated 2 days ago· 0

CV Deep Features

image

features
dims
count
success

◄model▾►

◄layer_name►

◄input_size224►

◄scale0.0039►

◄swap_rbtrue►

◄poolingglobal average pool (embedding)►

◄mean_r123.675►

◄mean_g116.280►

◄mean_b103.530►

◄std_r0.229►

◄std_g0.224►

◄std_b0.225►

Categoryimage/CV/ml

Inputs (13)

Name	Type	Default	Description
image	NPARRAY,IMAGE		Frames to embed: an IMAGE batch gives one embedding row per frame; an NPARRAY (gray/BGR/BGRA) is a single frame. Accepts a ComfyUI IMAGE/MASK directly (frame 0 of a batch) or an NPARRAY. Arithmetic ops (add, multiply, etc.) process the full IMAGE batch when both inputs have the same batch size.
model	COMBO		Backbone .onnx from ComfyUI/models/onnx (e.g. squeezenet1.1-7.onnx).
layer_name	STRING		Layer to tap, e.g. squeezenet0_concat7 (the 'onnx_node!' classic-engine prefix is added automatically if needed). Blank = the model's final output - usually class LOGITS, which transfer worse than an interior bottleneck.
input_size	INT	22416–2048	Square net input in px (224 for the ImageNet backbones); frames are resized by blobFromImage.
scale	FLOAT	0.00390–1000000	Pixel scale factor applied after mean subtraction; 1/255 = 0.0039215... maps 0-255 to the 0-1 range ImageNet models expect.
swap_rb	BOOLEAN	true	Swap R and B: OpenCV frames are BGR, most ONNX backbones want RGB - leave on.
pooling	COMBO	global average pool (embedding)	How a (C, H, W) activation map becomes one row: average each channel over space (C dims, position-invariant - the standard embedding), or flatten everything (CHW dims, keeps layout).
mean_ropt	FLOAT	123.675-255–255	RED channel mean subtracted before scaling. Default = ImageNet 0.485 * 255. Set 0 to disable.
mean_gopt	FLOAT	116.280-255–255	GREEN channel mean (ImageNet 0.456 * 255).
mean_bopt	FLOAT	103.530-255–255	BLUE channel mean (ImageNet 0.406 * 255).
std_ropt	FLOAT	0.2290.000001–255	RED channel std the scaled pixels are divided by (ImageNet 0.229). Set 1 to disable.
std_gopt	FLOAT	0.2240.000001–255	GREEN channel std (ImageNet 0.224). 1 disables.
std_bopt	FLOAT	0.2250.000001–255	BLUE channel std (ImageNet 0.225). 1 disables.

Outputs (4)

Name	Type	Description
features	NPARRAY	(N, D) float32 embeddings, one row per frame - feeds 'CV Train Classifier' / 'CV Stack Feature Classes'. Empty (0, 0) on failure.
dims	INT	Embedding length D (512 for SqueezeNet's squeezenet0_concat7).
count	INT	Frames embedded.
success	BOOLEAN	False when the layer name does not exist in the model or no frame could be embedded - gate downstream training with if/else.