IC-1485Sketch Transformer's FID degrades from 31.49 to 404.01 under Gaussian noise at the highest tested intensity

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

The paper evaluates the released Sketch Transformer on the photo-to-sketch task using 338 test photos from the CUFS dataset. Under five noise types at six intensity levels, FID degrades monotonically with noise intensity. For Gaussian noise, FID rises from 31.49 (clean) through 53.33, 69.88, 108.26, 124.72, 228.02, to 404.01 at s6. LPIPS similarly rises from 0.3055 to 0.6525. The degradation is consistent across all five noise types tested.

Evidence

correlational

Key metric

FID: 31.49 (clean) to 404.01 (Gaussian s6); LPIPS: 0.3055 (clean) to 0.6525 (Gaussian s6); 5 noise types x 6 intensity levels

Model

Sketch Transformer

Concepts

Failure mode

Datasets

CUFS [eval]

Methods

FID [eval], LPIPS [eval]

Related findings

IC-1484, IC-1486, IC-1487

Extraction

automatic-extraction