Like, the image i use to test the model has to be segmented also? Or i just can send in any picture i find online?
Can i somehow get a multilabel classification on a picture out of a segmentation model?
If you are training it with image + segmented image (mask), then yes for prediction you will need the same. If you are training an image to give the segmentation as the output, normal image is fine.
One option is have 2 models, one train it to generate segmentation from image and then another train it to detect fashion details from segmentation. That way, you can input a normal image, get the segmentation from your first model and finally get details from second. Another option is have a single model output both things (a bit harder) and another is to not use segmented information.