CIG‐Diff: Clinical‐Intent and Geometry‐Aware Diffusion for Low‐Light Endoscopic Image Enhancement

Low‐light image enhancement remains challenging due to the complex coupling between spatially varying illumination and surface reflectance. Existing approaches predominantly emphasize brightness correction or feature‐level reconstruction, often leading to color distortion and structural inconsistency under severe illumination degradation. To address these limitations, we propose CIG‐Diff, a geometry‐aware conditional diffusion framework for structure‐constrained low‐light enhancement, with endoscopic imaging as a representative application. Instead of treating enhancement as purely photometric correction, CIG‐Diff formulates it as a geometry‐consistent reconstruction process within a diffusion paradigm. We introduce a cross‐modal modulation (CMM) mechanism that incorporates CLIP‐encoded textual embeddings as high‐level task priors. These embeddings are injected into the diffusion U‐Net via conditional modulation to stabilize the denoising trajectory and maintain contextual coherence during restoration. To explicitly model illumination degradation, we further design a geometry‐aware illumination module (GAIM) inspired by Retinex theory. GAIM disentangles illumination and reflectance components and performs reflectance recalibration to mitigate spatially non‐uniform lighting while preserving geometric consistency and suppressing noise amplification. Extensive experiments demonstrate consistent improvements in quantitative performance and structural fidelity compared with state‐of‐the‐art low‐light enhancement methods.