GRDFNet is a lightweight image restoration network that combines gated and dilated residual blocks to deliver strong perceptual quality with modest compute requirements.
Recommended Configurations
num_sets = 3, feature_channels = 32: strong quality while staying fast for most desktop workloads.
num_sets = 6, feature_channels = 48: highest quality configuration; expect roughly a 4x slowdown versus the 32-channel model.
num_sets = 3, feature_channels = 24: suggested for lightly compressed video inference; typically 50~75% faster than the 32-channel variant when deployed with TensorRT.
Performance Snapshot
Example TensorRT run on an NVIDIA RTX 4080 Super (16 GB):
