"No one is harder on a talented person than the person themselves" - Linda Wilkinson ; "Trust your guts and don't follow the herd" ; "Validate direction not destination" ;

February 18, 2022

Hair styles research paper reads

Paper #1 - MichiGAN: Multi-Input-Conditioned Hair Image Generation for Portrait Editing

Code - Link

Key Notes

  • MichiGAN is capable of enabling multiple input conditions for disentangled hair manipulation.
  • Editing appearance (b), structure (c), and shape (d) while keeping the background unchanged
  • Disentangle the information of hair into a quartet of attributes – shape, structure, appearance, and background, and design deliberate representations
  • Appearance is encoded through our mask-transformed feature extracting network
  • Background encoder is placed parallel to the generation branch, which keeps background intact
  • An explicit disentanglement of hair visual attributes, and a
  • set of condition modules that implement the effective condition mechanism for each attribute with respect to its particular visual characteristics;
  • An end-to-end conditional hair generation network that provides complete and orthogonal control over all attributes individually or jointly;
  • An interactive hair editing system that enables straightforward and flexible hair manipulation through intuitive user inputs



  • We represent the hair shape as the 2D binary mask of its occupied image region
  • Backbone generation network to bootstrap the generator with specific appearance styles instead of random noises.
  • Force the GAN to reconstruct the same background content;

Loss Types - 

  • Feature matching loss. To achieve more robust training of GAN, we also adopt the discriminator feature matching loss
  • Perceptual loss. We also measure high-level feature loss with the pre-trained VGG19 model
  • Structural loss. We propose an additional structure loss to enforce the structural supervision

Paper #2 - LOHO: Latent Optimization of Hairstyles via Orthogonalization

Code is available at Link

Notes

  • Our approach decomposes hair into three attributes: perceptual structure, appearance, and style, and includes tailored losses to model each of these attributes independently.
  • Optimizing StyleGANv2’s extended latent space and noise space
  • Novel approach to perform hairstyle transfer on in-the-wild portrait images and compute the Frechet Inception Distance (FID) score. FID is used to evaluate generative models by calculating the distance between Inception [29] features for real and synthesized images in the same domain


  • Pretrained VGG [28] to extract high-level features 

Paper #3 - Applications of Generative Adversarial Networks in Hairstyle Transfer

Notes

  • InterFaceGan, StyleGan

Paper #4 - Learning to Generate and Edit Hairstyles

Notes

  • GAN model termed Hairstyle GAN (H-GAN)
  • Recognition, generation and modication of hairstyles, by using a single model.
  • VAEGAN [14] integrates the Variational Auto-Encoders (VAE) into GAN
  • InfoGAN [5] further models the noise variable z in Eq (1) by decomposing it into a latent representationy and incompressible noise z



Keep Exploring!!!

No comments: