|
--- |
|
datasets: |
|
- https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/ |
|
--- |
|
|
|
Scaled up version of https://huggingface.co/PrakhAI/AIPlane. This model generates 128x128 images of Aircrafts. |
|
|
|
It is trained on [FGVC Aircraft](https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/) images. Just like https://huggingface.co/PrakhAI/AIPlane, it uses [Progressive Growing](https://arxiv.org/pdf/1710.10196.pdf) with [Spectral Normalization](https://arxiv.org/pdf/1802.05957.pdf). |
|
|
|
The colab implementation can be found [here](https://colab.research.google.com/github/prakharbanga/AIPlane2/blob/main/AIPlane_ProGAN_%2B_Spectral_Norm_(128x128).ipynb). |
|
|
|
| Generated Images | Real Images (for comparison) | |
|
| -------- | --------- | |
|
|  |  | |
|
|
|
# Demo |
|
The demo app for this model is at https://huggingface.co/spaces/PrakhAI/AIPlane2 (please "Restart this Space" if prompted). |
|
|
|
The demo showcases random image generation and generation of images similar to a previously generated image. |
|
|
|
# Overview |
|
The model is able to produce a wide variety of images. This is likely owing to the use of [minibatch discrimination](https://arxiv.org/abs/1606.03498v1). |
|
|
|
The quality of the generated images is reminiscent of the early days of GANs. |
|
|
|
## Left-facing planes |
|
 |
|
|
|
## Right-facing planes |
|
 |
|
|
|
## Planes on the ground |
|
 |
|
|
|
## Poor quality images |
|
 |
|
|
|
## Unique generations |
|
 |
|
|
|
# Image memorization |
|
To check whether the Generator is simply memorizing particular images, latent space exploration is used. |
|
|
|
The latent vector for a generated image is slightly mutated, and the new images are compared to the original. The comparisons (shown below) suggest that mutations in the latent space translate to small (sometimes imperceptible) mutations in the generated images. |
|
|
|
These observations support the idea that the generator's space is continuous, and the generator is not restricted to a finite number of modes. |
|
|
|
| Original Generated Image | Neighbors in latent space | Diff against original image | |
|
| ----------- | -------- | ----------- | |
|
|  |  |  | |
|
|  |  |  | |
|
|  |  |  | |
