SRVGGNetCompact

A compact VGG-style network structure for super-resolution

Overview

SRVGGNetCompact is a compact VGG-style network architecture designed for efficient super-resolution. It performs upsampling in the last layer and conducts no convolution on the HR feature space, making it computationally efficient.

This architecture is used in the lightweight Real-ESRGAN models like realesr-animevideov3 and realesr-general-x4v3.

Class Definition

from realesrgan.archs.srvgg_arch import SRVGGNetCompact

model = SRVGGNetCompact(
    num_in_ch=3,
    num_out_ch=3,
    num_feat=64,
    num_conv=16,
    upscale=4,
    act_type='prelu'
)

Parameters

num_in_ch · int · default: "3"

Number of input channels. Typically 3 for RGB images.

num_out_ch · int · default: "3"

Number of output channels. Typically 3 for RGB images.

num_feat · int · default: "64"

Number of feature channels in intermediate layers. Higher values increase model capacity but also computational cost.

num_conv · int · default: "16"

Number of convolutional layers in the body network. More layers allow the model to learn more complex patterns.

upscale · int · default: "4"

Upsampling factor for super-resolution. Common values are 2, 4, or 8.

act_type · str · default: "prelu"

Activation function type. Options:

• 'relu': ReLU activation
• 'prelu': Parametric ReLU (default, learns activation parameters)
• 'leakyrelu': Leaky ReLU with negative slope of 0.1

Architecture Details

The network consists of:

1. Initial convolution: 3×3 conv layer that expands input channels to num_feat channels
2. Body network: num_conv layers of 3×3 convolutions with activation functions
3. Final convolution: Maps features to output space (channels = num_out_ch × upscale²)
4. Pixel shuffle upsampler: Rearranges feature maps to produce high-resolution output
5. Residual connection: Adds nearest-neighbor upsampled input to the network output

The network learns residual information rather than the full high-resolution image, which helps with training stability and performance.

Model Configurations

realesr-animevideov3 (XS size)

model = SRVGGNetCompact(
    num_in_ch=3,
    num_out_ch=3,
    num_feat=64,
    num_conv=16,  # Compact configuration
    upscale=4,
    act_type='prelu'
)

Optimized for anime video upscaling with minimal parameters.

realesr-general-x4v3 (S size)

model = SRVGGNetCompact(
    num_in_ch=3,
    num_out_ch=3,
    num_feat=64,
    num_conv=32,  # Deeper configuration
    upscale=4,
    act_type='prelu'
)

General-purpose model with more layers for better quality on diverse images.

Usage Example

import torch
from realesrgan.archs.srvgg_arch import SRVGGNetCompact

# Initialize model
model = SRVGGNetCompact(
    num_in_ch=3,
    num_out_ch=3,
    num_feat=64,
    num_conv=16,
    upscale=4,
    act_type='prelu'
)

# Load pretrained weights
model.load_state_dict(torch.load('realesr-animevideov3.pth')['params'])
model.eval()

# Inference
with torch.no_grad():
    lr_image = torch.randn(1, 3, 64, 64)  # Low-resolution input
    sr_image = model(lr_image)  # Output: (1, 3, 256, 256)

SRVGGNetCompact is significantly more efficient than RRDBNet, making it ideal for real-time applications and video processing.

Forward Method

def forward(x):
    """
    Args:
        x (Tensor): Input low-resolution image tensor of shape (B, C, H, W)
    
    Returns:
        Tensor: Super-resolved image of shape (B, C, H*upscale, W*upscale)
    """

The forward pass:

1. Processes input through body network layers sequentially
2. Applies pixel shuffle to upsample feature maps
3. Adds nearest-neighbor upsampled input as residual

Source

Defined in realesrgan/archs/srvgg_arch.py