tea_model.py

import torch
from torch import nn

# VAE Decoder but with half-size output.
# The last upsample is not there.

###
# Code from madebyollin/taesd

class Recon(nn.Module):
    def __init__(self, ch_in, ch_out):
        super().__init__()
        self.long = nn.Sequential(
            nn.Conv2d(ch_in, ch_out, 3, padding=1),
            nn.ReLU(),
            nn.Conv2d(ch_out, ch_out, 3, padding=1),
            nn.ReLU(),
            nn.Conv2d(ch_out, ch_out, 3, padding=1)
        )
        if ch_in != ch_out:
            self.short = nn.Conv2d(ch_in, ch_out, 1, bias=False)
        else:
            # The one without identity, a placeholder.
            self.short = nn.Identity()
        self.fuse = nn.ReLU()

    def forward(self, x):
        return self.fuse(self.long(x) + self.short(x))

class TeaDecoder(nn.Module):
    def __init__(self, ch_in):
        super().__init__()
        self.block_in = nn.Sequential(
            nn.Conv2d(ch_in, 64, 3, padding=1),
            nn.ReLU()
        )
        self.middle = nn.Sequential(
            *[Recon(64, 64) for _ in range(3)],
            # Opposite of stride=2
            nn.Upsample(scale_factor=2),
            # It leads to a simpler model with fewer parameters.
            # The output of the previous layers matches the number of channels specified in this line.
            # The input to this layer is already well-represented by the feature maps from the previous layers,
            # the bias may not add significant value.
            nn.Conv2d(64, 64, 3, padding=1, bias=False),
            # Final upscale to 1/2 size of the image.
            *[Recon(64, 64) for _ in range(3)],
            nn.Upsample(scale_factor=2),
            nn.Conv2d(64, 64, 3, padding=1, bias=False),
        )
        self.block_out = nn.Sequential(
            Recon(64, 64),
            # Convert to RGB, regardless of the latent channels.
            nn.Conv2d(64, 3, 3, padding=1),
        )


    def forward(self, x):
        # Clamp the input values to a specific range.
        clamped = torch.tanh(x / 1)
        cooked = self.middle(self.block_in(clamped))

        return self.block_out(cooked)