Pytorch implementation for the paper "Structural-analogy from a Single Image Pair".
Abstract: The task of unsupervised image-to-image translation has seen substantial advancements in recent years through the use of deep neural networks. Typically, the proposed solutions learn the characterizing distribution of two large, unpaired collections of images, and are able to alter the appearance of a given image, while keeping its geometry intact. In this paper, we explore the capabilities of neural networks to understand image structure given only a single pair of images, A and B. We seek to generate images that are structurally aligned: that is, to generate an image that keeps the appearance and style of B, but has a structural arrangement that corresponds to A. The key idea is to map between image patches at different scales. This enables controlling the granularity at which analogies are produced, which determines the conceptual distinction between style and content. In addition to structural alignment, our method can be used to generate high quality imagery in other conditional generation tasks utilizing images A and B only: guided image synthesis, style and texture transfer, text translation as well as video translation.
Python 3.7, Pytorch 1.4.0, argparse, Pillow 7.0.0, Scipy 1.4.1, skimage 0.16.2, numpy
You can train using the following command:
python train.py --input_a ./images/208.jpg --input_b ./images/209.jpg --gpu_id 0 --out ./output0/ --beta 10.0 --alpha 1.0
For other images, just replace input_a and input_b.
In many cases it is possible to improve results quality using one of the following hyperparameter change:
--beta 1.0
--min_size 25
--nfc 128 --min_nfc 128
--load checkpoint_folder
Where checkpoint_folder is the output folder of the run you wish to resume.
In some cases, the quality of the result can be improved using refinement. You can refine your results using SinGAN in the following way: Let "ab.png" the image that we want to refine using the original image "b.png" (i.e. "ab.png" should have the same patch distribution as "b.png"). First clone SinGAN repository
git clone https://github.com/tamarott/SinGAN
cd SinGAN
Then train a SinGAN network with "b.png" as input
python main_train.py --input_dir ../images/ --input_name b.png
You can refine using the command:
python paint2image.py --input_dir ./images/ --input_name b.png --ref_dir ./images/ --ref_name ab.png --paint_start_scale 4
Where paint_start_scale is a hyperparameter, and it is recommended to try several values (usually 3-5).
More details about SinGAN implementation can be found at the repository.
python train.py --input_a ./309.png --input_b ./308.png --gpu_id 7 --out ./output5/ --beta 10.0 --alpha 1.0 --lr_d 0.0005 --lr_g 0.0001 --lambda_g 0.1
Note that input_b should be the sketch. In many cases, refinement is useful (see the above instructions).
python train.py --input_a ./images/108.png --input_b ./images/109.png --gpu_id 0 --out ./output3/ --beta 10.0 --alpha 1.0 --min_size 25
python train.py --input_a ./images/10.png --input_b ./images/11.png --gpu_id 0 --out ./output4/ --beta 10.0 --alpha 1.0
python train.py --input_a ./images/8.png --input_b ./images/9.png --gpu_id 0 --out ./output2/ --beta 10.0 --alpha 1.0
cd video
Video to images:
python video2imgs.py --input ../video_data/volcano_video.mp4 --out ./volcano_imgs/
In some cases, quantization of the images helps to achieve better results (not true for most cases). For quantization:
python quant_vid.py --root_dir ./volcano_imgs/ --out ./volcano_q0/ --quant_level 2
Where quant_level control the level of quantization (higher increase in the number of colors). Training:
cd -
python train_video.py --video_dir ./video/volcano_q0/ --num_images 250 --input_b ./images/10.png --gpu_id 6 --out ./video/vid_out/ --vid_ext .png --min_size 25 --niter 25000
Evaluation:
python eval_video.py --video_dir ./video/volcano_q0/ --num_images 250 --gpu_id 6 --out ./video/eval_vid_out/ --load ./video/vid_out/ --min_size 25 --vid_ext '.png' --add_inject True --inject_level 8 --a2b True
Images to video:
python3 imgs2video.py --input ./eval_vid_out/ --out ~/Downloads/vid_out.avi --fps 25.0 --frames 250
Usually, evaluation is not necessary! because the training code already produces the desired output. If one uses a larger min_size argument the result would be aligned to fake samples which will be different from the real output, in this case it is useful to evaluate the real image. Evaluation is better when we inject the real image into one of the scales and then refine it using the generators before translating to the other domain. You can use the following command:
python eval.py --input_a ./208.jpg --input_b ./209.jpg --gpu_id 0 --out ./eval_output0/ --load ./output0/ --min_size 18 --img_size 220
Make sure to use the same min_size and img_size as in the training phase.
If you found this work useful, please cite.
@article{structuralanalogy2020,
author = {Sagie Benaim and
Ron Mokady and
Amit Bermano and
Daniel Cohen-Or and
Lior Wolf},
title = {Structural-analogy from a Single Image Pair},
journal = {Computer Graphics Forum},
volume = {n/a},
year = {2020},
doi = {https://doi.org/10.1111/cgf.14186},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.14186},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1111/cgf.14186}
}
For further questions, ron.mokady@gmail.com or sagiebenaim@gmail.com.
This implementation is heavily based on https://github.com/tamarott/SinGAN.