For the S-well series in the N Block of the Daqing Oilfield, well-logging curves commonly exhibit amplitude shifts, heterogeneous noise morphologies, and distribution drifts under multi-well and multi-interval conditions. As a result, conventional threshold-based cross-plots, feature engineering pipelines, and standard deep networks often fail to simultaneously achieve cross-well consistency, stratigraphic-boundary sensitivity, and physical interpretability. To address this challenge, we propose GeoDiff-Former, a unified-interpretation framework that integrates “curve normalization—stratigraphic representation—multi-task interpretation” into an end-to-end jointly optimized workflow. The method first introduces a conditional diffusion-based normalization module that models non-geological noise and acquisition-related drift as a learnable and reversible generative process, enabling task-driven adaptive distribution alignment and preventing acquisition effects from being mistaken as geological variations. It then constructs a geology-biased Transformer encoder, where relative-depth bias explicitly injects stratigraphic continuity and boundary discontinuity into the attention computation, strengthening stable characterization of thin interbeds and boundary-dominated intervals. Finally, a multi-task prediction head jointly performs facies classification, porosity inversion, and water/engineering-indicator discrimination, while a rock-physics consistency constraint is incorporated to suppress non-interpretable solutions and improve the reliability and practical utility of the outputs. The results demonstrate that GeoDiff-Former achieves more robust cross-well transfer and more coherent interval-wise interpretation in complex reservoirs, providing an innovative yet deployable deep-learning pathway for intelligent unified interpretation of well-logging data.