Spatially Resolved Transcriptomics (SRT) profiles gene expression with spatial context and advances studies of tissue organization and developmental processes. However, sparsity and complex noise introduced by technical limitations restrict the reliability of embeddings and obscure spatial domain identification. Although graph-based and dual-channel approaches attempt to integrate spatial and expression data, they still struggle under heterogeneous noise and often fail to recover biological structures. We design BioLemons to address these difficulties with a two-stage generative design. In the first stage, a Variational Autoencoder (VAE) extracts latent embeddings with graph neural backbones, which retain transcriptional relationships and neighborhood dependencies. In the second stage, a conditional diffusion process reconstructs expression with high fidelity. Graph-structured priors generate scale and shift information that modulate U-Net layers through affine transformations, while a mask-based augmentation strategy introduces diverse corruptions during training and improves robustness across noise conditions. Extensive experiments on both simulated and real datasets demonstrate the effectiveness of BioLemons. On the DLPFC dataset, it raises ARI by over 0.1 on the sliced 151673 and restores cortical boundaries across other slices. These results confirm that BioLemons reduces noise and produces spatial embeddings that align with biologically meaningful structures. The source code are available at  https://github.com/15831959673/BioLemons .