LTE

This repository contains the official implementation for LTE introduced in the following paper:

Local Texture Estimator for Implicit Representation Function (CVPR 2022)

Quick Start

1. Download a pre-trained model.

Model	Download
EDSR-baseline-LTE	Google Drive
EDSR-baseline-LTE+	Google Drive
RDN-LTE	Google Drive
SwinIR-LTE	Google Drive

2. Reproduce Experiments

Table 1: EDSR-baseline-LTE: bash ./scripts/test-div2k.sh ./save/edsr-baseline-lte.pth 0

Table 1: RDN-LTE: bash ./scripts/test-div2k.sh ./save/rdn-lte.pth 0

Table 1: SwinIR-LTE: bash ./scripts/test-div2k-swin.sh ./save/swinir-lte.pth 8 0

Table 2: RDN-LTE: bash ./scripts/test-benchmark.sh ./save/rdn-lte.pth 0

Table 2: SwinIR-LTE: bash ./scripts/test-benchmark-swin.sh ./save/swinir-lte.pth 8 0

Train & Test

[EDSR-baseline-LTE]

Train: python train.py --config configs/train-div2k/train_edsr-baseline-lte.yaml --gpu 0

Test: python test.py --config configs/test/test-div2k-2.yaml --model save/_train_edsr-baseline-lte/epoch-last.pth --gpu 0

[EDSR-baseline-LTE+]

Train: python train.py --config configs/train-div2k/train_edsr-baseline-lte-fast.yaml --gpu 0

Test: python test.py --config configs/test/test-fast-div2k-2.yaml --fast True --model save/_train_edsr-baseline-lte-fast/epoch-last.pth --gpu 0

[RDN-LTE]

Train: python train.py --config configs/train-div2k/rdn-lte.yaml --gpu 0,1

Test: python test.py --config configs/test/test-div2k-2.yaml --model save/_train_rdn-lte/epoch-last.pth --gpu 0

[SwinIR-LTE]

Train: python train.py --config configs/train-div2k/swinir-lte.yaml --gpu 0,1,2,3

Test: python test.py --config configs/test/test-div2k-2.yaml --model save/_train_swinir-lte/epoch-last.pth --window 8 --gpu 0

Model	Training time (# GPU)
EDSR-baseline-LTE	21h (1 GPU)
RDN-LTE	82h (2 GPU)
SwinIR-LTE	75h (4 GPU)

We use NVIDIA RTX 3090 24GB for training.