User:Aitantv/GAN: Difference between revisions
(→Intro) |
|||
| (3 intermediate revisions by the same user not shown) | |||
| Line 6: | Line 6: | ||
* First introduced in 2014 | * First introduced in 2014 | ||
* [https://thispersondoesnotexist.com thispersondoesnotexist.com] created by nvidia | * [https://thispersondoesnotexist.com thispersondoesnotexist.com] created by nvidia | ||
* The generator tries to create random synthetic outputs (for instance, images of faces), while the discriminator tries to tell these apart from real outputs (say, a database of celebrities). The hope is that as the two networks face off, they'll both get better and better—with the end result being a generator network that produces realistic outputs. | * Generator v Discriminator: The generator tries to create random synthetic outputs (for instance, images of faces), while the discriminator tries to tell these apart from real outputs (say, a database of celebrities). The hope is that as the two networks face off, they'll both get better and better—with the end result being a generator network that produces realistic outputs. Neither side of the equation should have the upper hand. | ||
Deep Learning Software | Deep Learning Software | ||
| Line 49: | Line 49: | ||
{{youtube|G3anJ03BPas}} | {{youtube|G3anJ03BPas}} | ||
== StyleGAN: Playing with Latent Space == | |||
{{youtube|dCKbRCUyop8}} | |||
Progressive Growing | |||
* Start with low res images & progresses to higher levels. Can take up to 10 days to get a convincing result. | |||
Latest revision as of 23:03, 7 December 2021
Intro
GANs: Generative Adversarial Networks
- First introduced in 2014
- thispersondoesnotexist.com created by nvidia
- Generator v Discriminator: The generator tries to create random synthetic outputs (for instance, images of faces), while the discriminator tries to tell these apart from real outputs (say, a database of celebrities). The hope is that as the two networks face off, they'll both get better and better—with the end result being a generator network that produces realistic outputs. Neither side of the equation should have the upper hand.
Deep Learning Software
- pix2pix
- StyleGan
Method
Generative Adversarial Networks (GAN)
- thispersondoesnotexist.com
- requires deep learning and neural network experience + coding experience
- GAN made up of two neural networks- Generator + Discriminator
- Introduced in 2013
- Generator - creates data that is preceived to be real. It recieves input and generates realistic images based on those images.
- Discriminator - decides which images created by the Generator are real and fake.
- StyleGan / RunwayML (web software/app) very easy to use / BigBiGan
DeepFake
- Deep Nostalgia
- Tokkingheads - combine a still with a video to make it talk
StyleGan2
- Create mega data set - at least 1000 images. You can use Fatkun plugin on chrome of DownThemAll to download all images on google.
- An algoithm can then auto-crop the data set so it's ready for the machine.
- Through 'transfer learning' you can first train the machine using one data set (e.g. Umbrellas), then add a second data set (e.g. Clouds), and it will project the first onto the second.
- Expect a final image/moving image of 512x512 or 1024x1024. You can always upscale to higher or lower resolutions
Voice
- You can also teach a machine a voice. Once it knows, you can replicate using text.
StyleGan
Text to Images
- Can create new bird species
- Incredibly detailed hi res image generation
- Shapeshifting
- 'Latent space'- the grey area where people dont really understand what's going on
Clips
StyleGAN: Playing with Latent Space
Progressive Growing
- Start with low res images & progresses to higher levels. Can take up to 10 days to get a convincing result.