alex-damian/pulse
Pulse
Pulse is a face image super-resolution tool and self-supervised image enhancer. It functions as a generative model image upsampler and latent space optimization tool designed to increase photo resolution and recover image details.
The system differentiates itself by using latent space exploration and spherical constraints to find high-fidelity matches within a generative model. It employs geodesic distance measurement and spherical latent space optimization to regularize representations and maintain parameter radii during the recovery process.
The project covers facial image restoration through a pipeline of automated face alignment and image upsampling. This includes detecting facial landmarks to standardize orientation and using reconstruction loss calculations to guide the restoration of low-resolution images.
Features
- Image Resolution Reconstruction - Increases the resolution of low-quality photos by reconstructing missing facial details using generative models.
- Face Alignment Tools - Provides utilities to detect facial landmarks and format face patches for high-fidelity upscaling.
- Image-to-Latent Projections - Maps existing low-resolution images back into the generative model's latent space to find high-resolution matches.
- Latent Search - Searches the generative model's internal representation to find a high-resolution match for a low-resolution input.
- Latent Vector Optimizers - Optimizes latent vectors using spherical constraints and reconstruction loss to recover missing image details.
- Latent-Based Upsamplers - Increases photo resolution by exploring the latent space of generative models for high-fidelity matches.
- Image Super Resolution Models - Functions as a specialized system for increasing the resolution and recovering fine details of facial images.
- Real Image Inversion - Implements techniques to map real facial images back to GAN latent codes for high-resolution reconstruction.
- Face Alignment - Detects facial landmarks and aligns image features to ensure consistent orientation for the generative model.
- Facial Restoration - Reconstructs high-fidelity facial details and removes artifacts using specialized alignment and generative synthesis.
- Geodesic Distance Regularizations - Implements geodesic distance measurement to regularize latent representations during the image recovery process.
- Spherical Constraints - Uses spherical constraints to maintain parameter radii while optimizing latent vectors for high-fidelity image reconstruction.
- Self-Supervised Reconstruction - Improves image quality without labeled training pairs by using the generative model's architecture as a regularization prior.
- Pixel-Wise Reconstruction Losses - Calculates the pixel-wise difference between generated and reference images to guide the upsampling process.
- Spherical Distance Calculators - Uses spherical distance calculations to regularize and constrain latent representations during the optimization process.
- Spherical Constraints - Restricts parameter updates to a spherical surface to maintain the original radius of latent vectors.
