VQVAE is trained to reconstruct the image, so in theory it should contain all the information (both content and location) inside its embeddings.