Image Classification

Image Classification is the node that tells you what category an image belongs to - "tabby cat," "sports car," "mountain." It's the classic computer-vision task, and in a ComfyUI graph it's a quick way to label images, route them by content, or filter a batch by what's actually in the frame. Think of it as the keyword-driven cousin of the captioning nodes: instead of a full description, you get one label and a confidence score.

It wraps the image-classification pipeline from Hugging Face, defaulting to microsoft/resnet-50. ResNet-50 is the most-studied image classifier in existence - a 2015 architecture with 50 layers that's still the reference point for "can it tell a dog from a cat." The node converts your image's first frame to PIL, runs the classifier, and returns the top prediction.

The inputs and output

• image - the IMAGE tensor to classify.
• model_name - Hub model ID as a string, default microsoft/resnet-50. The classifier's label set is baked into the model - whatever it was trained on (ResNet-50 uses ImageNet's 1000 categories) is all it can say.

Two outputs: label (a STRING - the predicted class) and score (a FLOAT confidence, 0–1). Wire both into display nodes; the score tells you whether the answer means anything.

How to install it

Standard pack install - ComfyUI Manager (search "ComfyUI-Transformers") or:

cd ComfyUI/custom_nodes
git clone https://github.com/kadirnar/ComfyUI-Transformers

Restart ComfyUI. ResNet-50 is about 100MB on first download, and requirements.txt covers the transformers stack.

Where people get burned

The label set is the whole game, and it's the part beginners miss. ResNet-50 only knows ImageNet's 1000 categories - no NSFW detection, no art style, no "anime" class. Ask it something outside its vocabulary and you'll get the nearest ImageNet label with low confidence. If you need a category it doesn't have, the pack's zero-shot nodes (which take free-text labels) are the better fit. Otherwise, for "give me a confidence-scored label for this image," this node is simple, fast, and does exactly what it says on the tin.