IC-1487CycleGAN's FID degrades from 76.92 to 180.82 under Gaussian noise for horse-to-zebra translation

Chaohua Shi, Kexin Huang, Lu GAN, Hongqing Liu, Mingrui Zhu, Nannan Wang, Xinbo Gao

SourceOn the Analysis of GAN-based Image-to-Image Translation with Gaussian Noise Injection

In the limitations section, the paper applies noise injection to the released CycleGAN for the horse-to-zebra translation task. The baseline CycleGAN's FID rises from 76.92 (clean) to 180.82 (Gaussian noise at sigma_e^2=0.16). The authors note that the noise-injected CycleGAN struggles with clean inputs (FID 283.97) and produces visible distortions around the zebra's head and legs despite better FID scores on noisy inputs, concluding that GNI is not a straightforward plug-and-play solution for bidirectional i2i models.

Evidence

correlational

Key metric

FID: 76.92 (clean) to 180.82 (sigma_e^2=0.16); noise-injected FID on clean: 283.97

Caveat

The authors explicitly state that for bidirectional i2i translation models, GNI 'cannot be considered a straightforward plug-and-play solution' and that 'adaptations in network architectures and/or loss functions might be requisite to achieve desired results.'

Model

CycleGAN

Concepts

Failure mode

Datasets

ImageNet-C [eval]

Methods

FID [eval]

Related findings

IC-1484, IC-1485, IC-1486

Extraction

automatic-extraction