When it comes to video game modding, its Bethesda’s RPGs like Skyrim and Fallout 4 that come to mind. However, Generative adversarial networks (a kind of neural networks) can apparently emulate the gameplay of visuals PUBG in Fortnite.
Okay, let’s break that down. These neural networks called GANs are quite proficient at imitating a particular art or visual style. We already know there’s no shortage of talented people in the video gaming community. One such person is Chintan Trivedi, who managed to train CycleGANs and convert some Fortnite footage to look like PUBG. The gameplay remains the same, only the oversaturated, cartoony visuals of Fortnite are replaced by those of PUBG . Although this is still far-fetched, it could help the community develop a wide range of modding tools for pretty much any game that allows it.
Generative adversarial networks
Before we move onto the juicy bits, lets take a look at how these neural networks work. Generative adversarial networks or GANs are deep neural net architectures consisting of two nets, pitting one against the other, hence the name “adversarial”.
These networks have immense potential as they can be taught to mimic literally any set of data. Meaning after a bit of exposure to existing images, music, human speech and poetry, GANs can produce bizarrely similar examples. They have been called robot artists, and their output is impressive – poignant even.
Generative vs. Discriminative Algorithms
To get a clearer picture of how GANs work, it’d apt to understand the working of discriminative and generative algorithms. I will keep it simple enough, without getting into too technical. Discriminative algorithms, as the name suggests differentiate between various data types.
Discriminative algorithms can be trained to detect spammy emails by identifying the words used in these mails. Later on these words can be compared those present in the new emails. By taking the number of matches found into account, the algorithm generates the probability of the email being junk. So, to sum up discriminative algorithms can be basically used to detect a certain data type.
On the other hand, a generative algorithm can be used to generate a certain kind of data set. For example, PR agencies can use them to create newer email layouts to try and bypass the spam check by including a different combination of words to advertise their products.
How GANs Were Used To Emulate PUBG in Fortnite
The GANs take the Fortnite image as the input and turns it into a PUBG’ed output. A generative algorithm is paired with a discriminative algorithm. The generative algorithm takes in the Fortnite input and converts it to the visual style of PUBG, while the discriminative algorithm will try to verify if the image is actually from PUBG or a fake created by the generative algorithm.
This way both the algorithms are constantly learning. Just like forgers continuously improve their skills and produce more and more realistic forgeries, while the police try to keep up and identify the real ones from the fake. The same way these algorithms compete with each other and as a result both improve over time.
Now, in this little experiment a second pair of discriminative/generating algorithms were also used to reconstruct the fake produced by the generator, back to the original input image from Fortnite. However, a condition was set to prevent the generator from returning the input.This was done so that the output footage had meshes from Fortnite, but the textures of PUBG.
However, one problem faced by Chintan Trivedi was that he was only able to use 256×256 images due to GPU memory limitations. This made the output pixelated and blurry. Someone with more horsepower, possibly a Tesla V100 should be able to work with higher resolutions.
The Output: PUBnite or FortG, Fornite With PUBG Graphics
The images generated by CycleGAN after 12 hours of training are quite promising. The network was successfully able to convert the sky, trees and grass from Fortnite to those of PUBG. The bright world of Fortnite was converted into the realistic looking one from PUBG. It even learned to replace the health bar of Fortnite with the gun and ammo indicator of PUBG. However, the network couldn’t establish a relationship between the character models of the two games which is why they look fuzzy. Watch the entire footage below:
Overall, the cycleGANs were able to do an impressive job of emulating Fortnite in PUBG. Although with the present hardware, it’s not feasible, but in the future as GPUs get more and more powerful, modders should be able to develop graphics packs, allowing players to play a particular title in the art-style of pretty much any game.