Usually this is done in three steps. The first step is using a neural network to create a bounding box around the object, then generating vector embeddings of the object, and then using similarity search on vector embeddings.
The first step is accomplished by training a detection model to generate the bounding box around your object, this can usually be done by finetuning an already trained detection model. For this step the data you would need is all the images of the object you have with a bounding box created around it, the version of the object doesnt matter here.
The second step involves using a generalized image classification model thats been pretrained on generalized data (VGG, etc.) and a vector search engine/vector database. You would start by using the image classification model to generate vector embeddings (https://frankzliu.com/blog/understanding-neural-network-embe...) of all the different versions of the object. The more ground truth images you have, the better, but it doesn't require the same amount as training a classifier model. Once you have your versions of the object as embeddings, you would store them in a vector database (for example Milvus: https://github.com/milvus-io/milvus).
Now whenever you want to detect the object in an image you can run the image through the detection model to find the object in the image, then run the sliced out image of the object through the vector embedding model. With this vector embedding you can then perform a search in the vector database, and the closest results will most likely be the version of the object.
Hopefully this helps with the general rundown of how it would look like. Here is an example using Milvus and Towhee https://github.com/towhee-io/examples/tree/3a2207d67b10a246f....
Disclaimer: I am a part of those two open source projects.