Fidelity Preservation in General-Purpose Image Editing Models Under Degradation

Ascertain whether general-purpose image editing models that perform conditional image-to-image generation can faithfully preserve source content under degradation, or whether their strong generative priors inherently induce semantic drift, hallucinated details, and structural deviations, thereby clarifying their suitability for restoration tasks requiring high fidelity.

Background

General-purpose image editing models leverage large-scale generative priors and instruction-based control, enabling flexible semantic edits but raising concerns about fidelity when applied to restoration.

The authors explicitly state uncertainty about these models’ ability to preserve original content under degradation versus introducing artifacts due to generative priors, prompting empirical investigation of this fidelity-preservation question.

References

Since these models are not explicitly optimized for recovering degraded observations, it remains unclear whether they can faithfully preserve source content under degradation, or whether their strong generative priors instead lead to semantic drift, hallucinated details, or structural deviations.

— Can Nano Banana 2 Replace Traditional Image Restoration Models? An Evaluation of Its Performance on Image Restoration Tasks  (2604.03061 - Sun et al., 3 Apr 2026) in Section 2.2 (Image Editing Models)

An important question is whether variation across FLEX-MRI trajectories provides complementary information about image content that remains weakly constrained by the measurements.

— Stochastic Finite-Transition Learning and Exploration for Accelerated MRI Reconstruction  (2610.13095 - Mirza et al., 9 Oct 2026) in Section Discussion

The source of this issue remains unknown, however it is suspected that the issue comes from increasing the variability of illumination in the low-light conditions.

— VesperMUSE: Multi-Modal Fusion Vision Transformer for Image Semantic Segmentation Under Diverse Light Conditions  (2610.21990 - Balmez, 15 Oct 2026) in Chapter Conclusions, Section “Known Limitations,” subsection “Synthetic Low-Light Adaptation Algorithm Strong Degradation”