Neural Texture Synthesis with Guided Correspondence - PyTorch version

Notice: Still reorganizing and testing codes... Feel free to contact me if something not work.

Official PyTorch implementation of "Neural Texture Synthesis with Guided Correspondence" [CVPR 2023]. More result can be seen here.

Prerequisites

Python 3.7.10
PyTorch 1.9.0

Texture Expansion

Synthesize texture with arbitrary size from exemplar texture:

python texturegen.py --data_folder=./data/stationary --output_folder=./outputs/result_synthesis --image_name=11.jpg --size=256 --output_size=512 512 --base_iters=500 --finetune_iters=0

Controlled Texture Synthesis

Synthesize texture based on progression, orientation, both or anotation:

1. Progression Control

python texturegen_aug.py --data_folder=./data/progression --output_folder=./outputs/result_progression --image_name=6.jpg --refer_prog_name=6_prog.png --trg_prog_name=1_prog.png --lambda_progression=50 --lambda_occurrence=0.05 --output_size=512 512 --scales=0.25 0.5 0.75 1 --use_flip

2. Orientation Control

python texturegen_aug.py --data_folder=./data/orientation --output_folder=./outputs/result_orientation --image_name=20.jpg --trg_orient_name=target_orient-1.npy --lambda_orientation=5 --lambda_occurrence=0.05 --output_size=512 512 --scales=0.25 0.5 0.75 1

3. Both Control

python texturegen_aug.py --data_folder=./data/both --output_folder=./outputs/result_both --image_name=1.jpg --refer_prog_name=1_prog.jpg --trg_prog_name=1_prog.jpg --trg_orient_name=x_orient.npy --lambda_progression=10 --lambda_orientation=1 --lambda_occurrence=0.05 --output_size=512 512 --scales=0.25 0.5 0.75 1

4. Label Control

python texturegen.py --data_folder=./data/label --output_folder=./outputs/result_label --image_name=1-image.jpg --refer_prog_name=1-label.jpg --trg_prog_name=cat.jpg --lambda_progression=10 --lambda_occurrence=0.05 --size=256 --output_size=512 512

Acknowledge

Sincerely thank xchhuang and tchambon for their wonderful works:

EliotChenKJ/Guided-Correspondence-Loss