RC500: https://drive.google.com/file/d/1x3tZmPw0IS9fxoKXel4xGT38nK8zd9hR/view?usp=sharing
We randomly selected 50 from 583 test images as the test set for comparative experiments:
test_dataset: https://drive.google.com/file/d/1Y3_4tuNcGbzj5bYr2JWhniVvm6W6J-Gb/view?usp=drive_link
Notably, we also selected 250 images from the 583 test images as the test set for ablation experiments to better verify the role of each component.
original_images: https://drive.google.com/file/d/1i9hv2yrG8cImo7In3KUiFnusn7GNC-7e/view?usp=drive_link
In this project, we use python 3.7.12 and pytorch 1.8.0, torchvision 0.9.0, cuda 11.1
We train the model using four GeForce RTX 3060.
bs = 8
lr= 0.0001
beta for EMA = (0.0, 0.99)
Supervised Pre-training max_steps=100000
Adversarial Training max_steps=20000
When there's a significant discrepancy between the color of the reference and that of the input image, it results in color distortion in the recolored image, causing unnaturalness.
When an unrealistic color palette is provided, the model generates semantically unreasonable images, such as recoloring trees to blue.
