taki0112/UGATIT
UGATIT
UGATIT is an unsupervised generative adversarial network and image-to-image translation model implemented in TensorFlow. It serves as the official research implementation of an ICLR 2020 paper, providing a framework for converting images between different visual styles without requiring paired training examples.
The system utilizes an unsupervised generative attentional network and attention maps to deform geometric shapes and modify textures during the translation process. It employs a cycle-consistent framework to ensure translation quality by requiring images to return to their original state after two-way domain shifts.
The codebase covers generative adversarial training and geometric image transformation, utilizing multi-scale discriminator architectures and adaptive layer-instance normalization to refine synthesis quality.
Features
- Unsupervised Style Translation - Provides a framework for unsupervised visual style translation by adjusting textures and geometric shapes without paired examples.
- Cycle Consistency Constraints - Ensures translation quality via cycle-consistency constraints that require images to be reconstructible after a two-way domain shift.
- Generative Adversarial Networks - Utilizes a generative adversarial network architecture to map images between two different domains.
- Image-to-Image Translation - Implements a TensorFlow-based image-to-image translation model for converting images between visual styles.
- Generative Adversarial Network Training - Employs a generative adversarial training workflow with competing generator and discriminator networks to synthesize realistic images.
- Attentional Generative Networks - Implements an unsupervised generative attentional network to translate images by attending to textures and geometric shapes.
- Unpaired Image Translation - Implements unpaired image translation to convert images between visual styles without requiring paired training data.
- Deformable Attention - Implements deformable attention layers to warp geometric shapes during the image translation process.
- Attentional Image Synthesis Tools - Ships a computer vision system that uses attention maps to deform geometric shapes and modify textures.
- Computer Vision Research - Provides the official research implementation of a computer vision model based on the ICLR 2020 paper.
- Cycle-Consistent Frameworks - Implements a cycle-consistent generative framework to ensure images return to their original state after two-way domain shifts.
- Multi-Scale Discriminators - Utilizes a multi-scale discriminator architecture to evaluate images at various resolutions for improved structural coherence.
- Image-to-Image Synthesis Frameworks - Implements a framework for image-to-image synthesis using deep learning to generate images based on learned visual patterns.
- Paper Implementations - Provides the official code implementation of the research described in an ICLR 2020 paper.
- Image Geometric Transformations - Performs geometric image transformations to handle non-rigid spatial adjustments and shape modifications during translation.
- Computer Vision Libraries - Unsupervised generative attentional networks for image translation.
