Writing Rectified Flow Network in Python Part 1 - The Autoencoder
This autoencoder will compress the size from 3x128x128 to 3x16x16. This smaller size is easier for the rectified flow network to learn.
The block of the autoencoder is based on MobileNetV2, it is basically an inverted residual block (the channels are expanded inside the block instead of compressed).
For the encoder I use strided convolution to retain as much information as possible. But on the decoder, I use nearest neighbor upsampling because it yields the best reconstruction.
To increase the sharpness of the output, I added vgg loss and laplacian loss. Then to prevent the latent space from exploding, I added a loss term that will push the latent space from -1 to 1.