Low-light image enhancement faces a fundamental trade-off. Generative models can recover plausible texture but risk hallucinating structures that are not actually present. Restoration networks, on the other hand, preserve the original structure but may leave some regions under-enhanced. The abstract notes that neither advantage is spatially uniform, meaning the best choice can vary across different parts of an image.
The paper addresses this by asking when a low-light pipeline should generate, and proposes a selective approach guided by auditable risk control. Rather than committing to one method globally, the framework appears to decide case-by-case whether to use generative enhancement or restoration, with the goal of keeping hallucination risk in check while still achieving adequate enhancement.
The significance of this work is in making the decision process auditable. Users can see and verify when and why generative enhancement is applied, which is a key step toward deploying such methods in sensitive imaging applications where trust and transparency matter.