Why generated image share very similar visual content with original image?

[whuzjw]

Great work! But I am a little confused about the high similarity between original and generated image. Because the input image is a random noise and there is no conditioning when inferencing, the generated image should be a random image instead of a very certain image (very similar to original image). I wonder if the reason is that the original image must be among the training images and the input noise when inferencing is similar to that when training. Thank you in advance for your help.