# vae_encoder.py import torch import torch.nn as nn from config import IMG_CHANNELS, LATENT_CHANNELS, VAE_BASE_CHANNELS, VAE_CHANNEL_MULT class ResBlock(nn.Module): """GroupNorm -> SiLU -> Conv, twice, plus a skip connection. The skip lets fine detail (thin text strokes!) pass straight through; the old plain conv stack had to squeeze everything through each layer.""" def __init__(self, in_ch, out_ch): super().__init__() self.block = nn.Sequential( nn.GroupNorm(32, in_ch), nn.SiLU(), nn.Conv2d(in_ch, out_ch, 3, padding=1), nn.GroupNorm(32, out_ch), nn.SiLU(), nn.Conv2d(out_ch, out_ch, 3, padding=1), ) self.skip = nn.Conv2d(in_ch, out_ch, 1) if in_ch != out_ch else nn.Identity() def forward(self, x): return self.skip(x) + self.block(x) class VAEEncoder(nn.Module): """ 256x256x3 image -> 32x32x4 latent (mu, log_var). One resolution level per entry in VAE_CHANNEL_MULT, downsampling between levels: 256 -> 128 -> 64 -> 32. """ def __init__(self, in_channels=IMG_CHANNELS, latent_channels=LATENT_CHANNELS, base_channels=VAE_BASE_CHANNELS, channel_mult=VAE_CHANNEL_MULT): super().__init__() chs = [base_channels * m for m in channel_mult] layers = [nn.Conv2d(in_channels, chs[0], 3, padding=1)] for i, ch in enumerate(chs): prev = chs[i - 1] if i > 0 else chs[0] layers += [ResBlock(prev, ch), ResBlock(ch, ch)] if i < len(chs) - 1: layers.append(nn.Conv2d(ch, ch, 3, stride=2, padding=1)) # downsample x2 layers += [ResBlock(chs[-1], chs[-1]), nn.GroupNorm(32, chs[-1]), nn.SiLU(), nn.Conv2d(chs[-1], 2 * latent_channels, 3, padding=1)] self.net = nn.Sequential(*layers) def forward(self, x): mu, log_var = self.net(x).chunk(2, dim=1) return mu, log_var.clamp(-30.0, 20.0) # keep exp(log_var) finite