• Perceptual loss of images generated by AutoEncoder
• Using the pre-trained AutoEncoder trained on ImageNet, encoded the images into embeddings on the latent space.
• Applied the vector difference of two images with different conditions like dry and wet, transparency, to the latent vector of the input image, and generated the new image
• Calculated the LPIPS(Learned Perceptual Image Patch Similarity) distance of images of the latent vectors added by different Gaussian noise, and compared with human evaluation
• Framework & Language: PyTorch, Python

dry                  wet

Using the difference of the encoded latent vectors in the AutoEncoder, we could apply the condition into other images.

• Below is the generated images with wet condition.

dry                  wet

dry                  wet

• With different parameters

• reverse condition

• The same idea, using the transpanrency diffrence, apply it into the object

• Recently, I'm studying the LPIPS distance (A perceptual metric of image similarity) of the generated image.

• Fancy geenrated image