Low-light image enhancement is vital for computer vision, yet existing methods like Zero-DCE have drawbacks: their fixed-step recursive enhancement often causes over-or insufficient enhancement, and ReLU’s non-smoothness leads to abrupt brightness transitions and gradient vanishing in extremely dark regions, limiting fine-detail recovery. To solve these issues, this paper proposes the AC-Net low-light enhancement network with two key innovations. First, a feature extraction module based on smooth nonlinear mapping replaces ReLU with Softplus—this improves illumination transition smoothness and mitigates gradient vanishing. Combined with a U-Net encoder–decoder structure and skip connections, it enables multi-scale feature fusion and high-quality image reconstruction. Second, an adaptive enhancement step-size module extracts a global illumination descriptor vector from shared features; a fully connected layer predicts enhancement strength ratios, and a probabilistic sampling mechanism maps these ratios to discrete steps (larger steps for darker areas to ensure sufficient detail enhancement, smaller steps for brighter regions to avoid overexposure). AC-Net produces enhanced images with smoother illumination, fewer artifacts/color distortions, and richer high-frequency details. Experiments on public datasets show it outperforms state-of-the-art methods in PSNR, SSIM, and LPIPS, with stability in cross-domain scenarios (e.g., UIEB underwater dataset). Ablation experiments further verify each module’s role in performance improvement.