Revealing the Secrets of Creating New Game Characters Without Using Paintbrush

  • Instability in training-Mode collapse is an inherent problem in the GAN generator, which occurs when the generator doesn’t learn any useful representation which is indistinguishable from the real images, thus making the discriminator always win.
  • Capable of generating relatively smaller images, such as 64*64 pixels.
  • These limitations are overcome by using a more sophisticated model, known as Progressive Growing GAN (ProGAN), specifically designed to minimize the limitations of vanilla GANs.
Fig 1. A- Random generation of MK characters using Random Noise vector as input, B-Controlled generation of MK characters using Random Noise vector and labeled data as input (Condition-Create MK characters with features as Male, Beard, Moustache)

Architectural details of StyleGANs

Following are the architectural changes in styleGANs generator:

a) ProGAN generator b) StyleGAN generator
  1. ‘config-a’, # Baseline Progressive GANs

Experimental setup

Table 2: Dataset summary

Fig 3. Sample images from the training MK dataset

Experimental design

We have conducted a set of experiments to examine the performance of StyleGANs in terms of FID, quality of output produced, training time vs performance on FID. In addition, we also checked the results imposing pre-trained latent vectors on new faces of data and Mortal Kombat characters data. We have implemented GANs and performed its feasibility analysis to overcome the following issues:

Evaluation metrics

FID [Heusel et al. 2017]-Frechet Inception Distance score (FID) is a metric for image generation quality that calculates the distance between feature vectors calculated for real and generated images. FID is used to understand the quality of the image generated, the lower the FID score, higher the quality of image generated. The perfect FID score is zero.

Experimental platform

All experiments are performed on the AWS platform, and the following is its configuration.

Experiments results

In order to provide a precise view of image generation of MK characters, we have conducted various experiments; and we were able to extract different results for each experiment.

Fig 4. MK training results 1-Kims 2364, 2-Kims 5306 , 3-Kims 6126, 4-,Kims 8006 5-Kims 9766, 6- 10000 Kims
Table 3. Time taken for training & FID score
Fig 5. Real-time images generation (using seed value) @7726 kims pickle MK.
Fig 6. Style Mixing results on MK characters
Fig 7. Progressive GANs output

Conclusion

In this whitepaper, we have discussed the methodology of applying the feasibility analysis using styleGANs for Mortal Kombat characters. We have provided a detailed report on the GANs types and evolution with respect to image generation approaches to solve different use cases using GANs. Which also includes style representation and conditional generation. The latter part of our attempt provides the results of GANs training time, FID scores, real-time image generation output, style mixing, and ProGANs results. After completing ~15 days of training (with GPU cost $1.14 per hour), the FID score achieved is 53.39. The quality of images is also improved, which can be further enhanced by conducting flawless training sessions. Recent advancements in GANs illustrate that a better GAN performance is achievable even with fewer data. Adaptive Discriminator Augmentation [Karras et al., 2020] and Differentiable Augmentation [Zhao et al., 2020] are a few of the recent approaches which have been proposed to train GANs effectively even with less amount of data, which is currently being researched in our CoE team.

About Affine

Affine is a Data Science & AI Service Provider, offering capabilities across the analytical value chain from data engineering to analytical modeling and business intelligence to solve strategic & day-to-day business challenges of organizations worldwide. Affine is a strategic analytics partner to medium and large-sized organizations (majorly Fortune 500 & Global 1000) around the globe that creates cutting-edge creative solutions for their business challenges.

References

[1] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., … & Bengio, Y. (2014). Generative adversarial nets. In Advances in neural information processing systems (pp. 2672–2680).

--

--

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Affine

Affine

Affine is a provider of analytics solutions, working with global organizations solving their strategic and day to day business problems www.affineanalytics.com