nprithviraj24 / face-hallucination Goto Github PK

Following models are part of my Masters' thesis (pdf)

1. CycleGAN-with-EDSR
2. Image-Degrade
   • To learn image super-resolution, use a GAN to learn how to do image degradation first
3. AdaIN for degradation
   • Based on Adaptive Instance normalisation from style transfer literature
4. degradeSR
   • Degrade an image using AdaIN and then perform super-resolution.

Map two different domains and generate high-quality images.

• CYCADA: Cycle consistent Adversarial Domain Adaptation

  • Adversarial adaptation models: Feature spaces discover domain invariant representation, but are different to visualise and sometimes fail to capture pixel-level and low-level domain shifts.

  • Alignment typically involves minimizing some measure of distance between the source and target distributions such as

     - Maximum mean discrepancy
     - Correlation distance
     - Adversarial discriminator accuracy
    

  This method is only for domain adaptation, however the conjecture is that we can do super-resolution along with domain adaptation.

• Attribute-Guided Face Generation Using Condition CycleGAN

• StyleVAE: Style basedVAE for Real-World SR StyleVAE + SR Network

** Yet to Come **

FIDs scores between model and interpolation technique

How can we evaluate our model if we don't have groud-truth high resolution images?

We can have a decent workaround for this problem, by calculating the FID of our generated images with frame of reference dataset such as Celeb-A, FFHQ, AIT3D etc. Idea is that we calculate FID between the interpolated high-res and frame of reference dataset. We compare this value with FID between synthesized high res image from our model and frame of reference dataset

Reference Dataset CelebA (from Kaggle)

*With official pretrained weights.