Illumination-specific superiority of prompt conditioning

Determine whether illumination-conditioned prompts are superior to learned fixed prompts in mean performance for low-light image enhancement, beyond the usefulness of the shared bottleneck-side sidecar pathway.

Background

The paper compares three matched variants of an SSM-based low-light image enhancement system: a Base restorer without the sidecar pathway, a Prompt-Adapter that generates input-dependent prompts from an estimated illumination map, and a Learned-Fixed control that uses a shared learned prompt bank. The Prompt-Adapter improves over the Base model, but the Learned-Fixed control achieves a slightly higher mean PSNR on LOL-v1 under the locked three-seed protocol.

Because both prompt-based variants provide comparable gains over the Base model, the reported evidence attributes the improvement more securely to the sidecar conditioning pathway than to illumination-specific, input-dependent prompting. The authors therefore leave unresolved whether illumination-conditioned prompts themselves provide a strict mean-performance advantage over learned fixed prompts.