Emerging imaging systems must handle large amounts of noisy data under strict resource constraints. Instead of denoising and compressing in two separate steps, this work integrates both tasks into a single framework based on the Mean Scale Hyperprior architecture. By incorporating denoising directly in the compression process, the model dedicates fewer bits to encoding noise and focuses on preserving essential image details. Two training modalities are explored: one where the model is trained to reconstruct noisy inputs, and another where it learns to produce clean outputs from noisy inputs. Experiments using SEN12MS satellite images and a custom noise model show that models trained to generate clean outputs achieve higher-quality reconstructions and better rate-distortion performance.