Results & Discussion
Animation
![](/766-Project-Public/assets/images/result2-e52e25e7d9de07f15b9089c1edf5a380.gif)
If you haven't seen this animation in our home page, this is the style transfer in action.Comparison with others
CycleGAN & DiscoGAN
![](/766-Project-Public/assets/images/cycleresult-c54c5df6838b88fef7164ebf4f318d0a.png)
Result of CycleGAN & DiscoGANOverall the characters generated by CycleGAN are recognizable, however, some characters suffered from missing/broken strokes and some other appeared in the wrong direction.
DualGAN
![](/766-Project-Public/assets/images/dualresult-a00de2e87d7ee73d9c1658be90ed5c3c.png)
Result of DualGANFor DualGAN, the structure of its generator and discriminator helps it to maintain a more stable and complete output structure. However, the strong constraints prevent it from learning the main characteristics of different fonts.
Ours
![](/766-Project-Public/assets/images/ouresult-0dd49cc3c3b99959adb9c518d6e8e957.png)
Result of oursThe results from our model is visibily the best across all 4 fonts as we elimated previous issues that appeared in CycleGAN.
Gallery
All models are trained for 100 epochs.
Style B
![](/766-Project-Public/assets/images/result1-553f567dd10e81fe5205ed8c48a2e4b4.png)
Style C
![](/766-Project-Public/assets/images/result2-616797544c933fc6d143e14a0637c977.png)
Style D
![](/766-Project-Public/assets/images/result3-53c458831dfca6f3472790f87580e131.png)
Style E
![](/766-Project-Public/assets/images/result4-6e480653b1ebd59ee760c84ad6de5caa.png)
Full Gallery of our model
(Require UW Madison Login)
Style B
Style C
Style D
Style E