Homework 1 (Color-Transfer and Texture-Transfer)
Training MUNIT
Under the limit of time and space constraint, we train MUNIT on summer2winter_yosemite dataset only with batch size 1 and 200000 iterations. All the images are trained and tested under resolution of 256x256.
The training loss is depicted as below figure:
| Discriminator Loss | Generator Loss |
|---|---|
 |
 |
Inference one image in multiple styl
Some result of the trained model are shown in below:
Summer to Winter
| Input (summer) | Output (winter) |
|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
Winter to Summer
| Input (winter) | Output (summer) |
|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
Result of each content style combination
| Summer Content Winter Style |
 |
 |
 |
|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
| Winter Content Summer Style |
|
|
|
|---|---|---|---|
|
 |
 |
 |
|
 |
 |
 |
|
 |
 |
 |
The result shown above is already convince us that the model is capable of transfering style and content between given two images. However, we do observe some failure result like:
| 
| 
| 
|
Some possible reasons:
- Our training time is not enough
- The random sampled style image is not suitable for the content
Compare with other method
We show some result of neural style transfer with ImageNet pretrained vgg19 directly applied on the dataset we use.
| Summer Content Winter Style |
|
|
|
|---|---|---|---|
|
 |
 |
 |
|
 |
 |
 |
|
 |
 |
 |
| Winter Content Summer Style |
|
|
|
|---|---|---|---|
|
 |
 |
 |
|
 |
 |
 |
|
 |
 |
 |
As the model is only pre-trained on ImageNet, the result is not as well as expected with artist painting like result. A clever choice for such pre-trianed setting is to use artist style to another arthis style transfer while realistic scene is not a good choice for such setting.