Skip to content
master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.

SinGAN

Project | Arxiv | CVF | Supplementary materials | Talk (ICCV`19)

Official pytorch implementation of the paper: "SinGAN: Learning a Generative Model from a Single Natural Image"

ICCV 2019 Best paper award (Marr prize)

Random samples from a single image

With SinGAN, you can train a generative model from a single natural image, and then generate random samples from the given image, for example:

SinGAN's applications

SinGAN can be also used for a line of image manipulation tasks, for example: This is done by injecting an image to the already trained model. See section 4 in our paper for more details.

Citation

If you use this code for your research, please cite our paper:

@inproceedings{rottshaham2019singan,
  title={SinGAN: Learning a Generative Model from a Single Natural Image},
  author={Rott Shaham, Tamar and Dekel, Tali and Michaeli, Tomer},
  booktitle={Computer Vision (ICCV), IEEE International Conference on},
  year={2019}
}

Code

Install dependencies

python -m pip install -r requirements.txt

This code was tested with python 3.6, torch 1.4

Please note: the code currently only supports torch 1.4 or earlier because of the optimization scheme.

For later torch versions, you may try this repository: https://github.com/kligvasser/SinGAN (results won't necessarily be identical to the official implementation).

Train

To train SinGAN model on your own image, put the desired training image under Input/Images, and run

python main_train.py --input_name <input_file_name>

This will also use the resulting trained model to generate random samples starting from the coarsest scale (n=0).

To run this code on a cpu machine, specify --not_cuda when calling main_train.py

Random samples

To generate random samples from any starting generation scale, please first train SinGAN model on the desired image (as described above), then run

python random_samples.py --input_name <training_image_file_name> --mode random_samples --gen_start_scale <generation start scale number>

pay attention: for using the full model, specify the generation start scale to be 0, to start the generation from the second scale, specify it to be 1, and so on.

Random samples of arbitrary sizes

To generate random samples of arbitrary sizes, please first train SinGAN model on the desired image (as described above), then run

python random_samples.py --input_name <training_image_file_name> --mode random_samples_arbitrary_sizes --scale_h <horizontal scaling factor> --scale_v <vertical scaling factor>

Animation from a single image

To generate short animation from a single image, run

python animation.py --input_name <input_file_name> 

This will automatically start a new training phase with noise padding mode.

Harmonization

To harmonize a pasted object into an image (See example in Fig. 13 in our paper), please first train SinGAN model on the desired background image (as described above), then save the naively pasted reference image and it's binary mask under "Input/Harmonization" (see saved images for an example). Run the command

python harmonization.py --input_name <training_image_file_name> --ref_name <naively_pasted_reference_image_file_name> --harmonization_start_scale <scale to inject>

Please note that different injection scale will produce different harmonization effects. The coarsest injection scale equals 1.

Editing

To edit an image, (See example in Fig. 12 in our paper), please first train SinGAN model on the desired non-edited image (as described above), then save the naive edit as a reference image under "Input/Editing" with a corresponding binary map (see saved images for an example). Run the command

python editing.py --input_name <training_image_file_name> --ref_name <edited_image_file_name> --editing_start_scale <scale to inject>

both the masked and unmasked output will be saved. Here as well, different injection scale will produce different editing effects. The coarsest injection scale equals 1.

Paint to Image

To transfer a paint into a realistic image (See example in Fig. 11 in our paper), please first train SinGAN model on the desired image (as described above), then save your paint under "Input/Paint", and run the command

python paint2image.py --input_name <training_image_file_name> --ref_name <paint_image_file_name> --paint_start_scale <scale to inject>

Here as well, different injection scale will produce different editing effects. The coarsest injection scale equals 1.

Advanced option: Specify quantization_flag to be True, to re-train only the injection level of the model, to get a on a color-quantized version of upsampled generated images from the previous scale. For some images, this might lead to more realistic results.

Super Resolution

To super resolve an image, please run:

python SR.py --input_name <LR_image_file_name>

This will automatically train a SinGAN model correspond to 4x upsampling factor (if not exist already). For different SR factors, please specify it using the parameter --sr_factor when calling the function. SinGAN's results on the BSD100 dataset can be download from the 'Downloads' folder.

Additional Data and Functions

Single Image Fréchet Inception Distance (SIFID score)

To calculate the SIFID between real images and their corresponding fake samples, please run:

python SIFID/sifid_score.py --path2real <real images path> --path2fake <fake images path> 

Make sure that each of the fake images file name is identical to its corresponding real image file name. Images should be saved in .jpg format.

Super Resolution Results

SinGAN's SR results on the BSD100 dataset can be download from the 'Downloads' folder.

User Study

The data used for the user study can be found in the Downloads folder.

real folder: 50 real images, randomly picked from the places database

fake_high_variance folder: random samples starting from n=N for each of the real images

fake_mid_variance folder: random samples starting from n=N-1 for each of the real images

For additional details please see section 3.1 in our paper