In silico labeling predicts organelle-specific localizations from label-free images, potentially enabling longitudinal and multiplexed live-cell imaging. Here we present DM4ISL (Diffusion Model for In Silico Labeling), a high-fidelity and uncertainty-aware model that overcomes current limitations, such as blurry or structurally inaccurate predictions, by resolving the intrinsic entanglement of biological signal and noise that characterizes fluorescence imaging. DM4ISL leverages the iterative nature of diffusion through a step-wise self-ensemble averaging approach that marginalizes these stochastic fluctuations, preventing the overfitting to noise that typically degrades standard final-step predictions. Benchmarking across six organelles demonstrates that DM4ISL systematically outperforms state-of-the-art architectures in structural fidelity, as confirmed through both pixel-based and application-specific metrics. Furthermore, calculating the standard deviation across intermediate denoising steps generates localized uncertainty maps that automatically detect erroneous predictions at inference time. Providing high-fidelity predictions alongside built-in reliability metrics, DM4ISL offers a robust method for trustworthy, data-driven discovery in cell biology.