Core Concepts
EGIC is a novel generative image compression method that efficiently traverses the distortion-perception curve, outperforming state-of-the-art methods.
Abstract
1. Introduction:
Neural image compression with generative models achieves high perceptual quality at low bit-rates.
The trade-off between distortion and perception is crucial in lossy compression.
2. Related Work:
Diffusion models challenge GANs in generative image compression.
Criticism of transparency in generation process addressed by various methods.
3. Background:
Traditional rate-distortion trade-off involves encoder, decoder, and entropy model.
Rate-distortion-perception trade-off navigated using discriminator D with non-saturating loss.
4. Our Approach:
EGIC introduces OASIS-C discriminator and ORP retrofit solution for multi-realism compression.
Two-stage training strategy fine-tunes generator while keeping encoder and pre-trained models fixed.
5. Exploring GANs for Compression:
Comparison of discriminator architectures like PatchGAN, SESAME, U-Net, Projected, and OASIS shows OASIS excels in perception.
6. Improving OASIS:
Weight normalization and projection improve OASIS performance step-by-step.
7. Comparison to the State-of-the-Art:
EGIC competes with diffusion and GAN-based methods on CLIC 2020 dataset, showing superior performance in low to medium bit-ranges.
Stats
EGICは、他の方法に比べて優れた性能を示す:FIDスコアが16.50でPSNRが30.03。