YoVDO

VQ-GAN - Taming Transformers for High-Resolution Image Synthesis - Paper Explained

Offered By: Aleksa Gordić - The AI Epiphany via YouTube

Tags

Generative Adversarial Networks (GAN) Courses Deep Learning Courses GPT-2 Courses Image Synthesis Courses

Course Description

Overview

Explore a comprehensive video explanation of the VQ-GAN (Vector Quantized Generative Adversarial Network) paper, focusing on high-resolution image synthesis using transformers. Dive into key modifications of VQ-VAE, including perceptual loss and adversarial loss for crisper outputs. Learn about sequence prediction with GPT, generating high-resolution images, and in-depth loss explanations. Discover transformer training techniques, conditioning methods, and various sampling strategies. Compare results with other models, including DALL-E, and understand the effects of receptive fields on image generation.

Syllabus

Intro
A high-level VQ-GAN overview
Perceptual loss
Patch-based adversarial loss
Sequence prediction via GPT
Generating high-res images
Loss explained in depth
Training the transformer
Conditioning transformer
Comparisons and results
Sampling strategies
Comparisons and results continued
Rejection sampling with ResNet or CLIP
Receptive field effects
Comparisons with DALL-E


Taught by

Aleksa Gordić - The AI Epiphany

Related Courses

Artificial Creativity
Parsons School of Design via Coursera
Building Language Models on AWS (Japanese)
Amazon Web Services via AWS Skill Builder
Deep Learning NLP: Training GPT-2 from scratch
Coursera Project Network via Coursera
Generating New Recipes using GPT-2
Coursera Project Network via Coursera
Accelerating High-Performance Machine Learning at Scale in Kubernetes
CNCF [Cloud Native Computing Foundation] via YouTube