The paper measures GP-UNIT's output quality under five noise types (Gaussian, uniform, color, Laplacian, salt-and-pepper) at six intensity levels. In the reference-guided setting, FID degrades from 17.82 (clean) to 36.84 (Gaussian s6), a clear failure mode. In the latent-guided setting, however, FID starts at 22.82 (clean) and actually decreases to 21.76 at Gaussian s6, with KID dropping from 9.67 to 7.88. The authors flag this as 'intriguing behavior' and state it will be the subject of future investigation, noting that GP-UNIT integrates an additional latent code or reference image that their theoretical framework does not account for.
Evidence
correlational
Key metric
Reference-guided FID: 17.82 (clean) to 36.84 (Gaussian s6); Latent-guided FID: 22.82 (clean) to 21.76 (Gaussian s6); Latent-guided KID: 9.67 (clean) to 7.88 (Gaussian s6)
Caveat
The authors note that the FID results for GP-UNIT in the unsupervised setting diverge from theoretical KL-divergence predictions, possibly because GP-UNIT integrates an additional latent code or reference image not contemplated by the theory. They state a comprehensive analysis of this behavior is left for future work.