Plastic injection molding requires frequent retuning of process parameters as temperature and humidity drift, yet setup still depends on skilled operators’ trial-and-error. We present a generative inference system that recommends multiple feasible process parameter sets for a desired quality under measured environmental conditions. A real injection molding testbed was instrumented to record four environmental variables (machine/factory temperature and humidity), ten key process parameters (three injection speeds, three injection pressures, three injection positions, and hold time), and good/defective labels. First, a surrogate model learns a mapping from (process parameters and environmental variables) to quality and serves as a virtual experiment for fast screening of candidate settings. Next, we train a classifier-free guidance diffusion model to sample the conditional distribution of process parameters for defect-free or defective outcomes, enabling stable training and broad distribution coverage. Compared with conditional GAN and conditional VAE, the diffusion model achieves substantially higher fidelity and diversity: only 1.63% of generated settings are misgenerated, versus 23.42% and 44.54%, respectively, and UMAP visualizations suggest reduced mode collapse and broader coverage of feasible defect-free settings. Finally, experimental validation on the injection molding testbed under multiple ambient conditions confirms practical utility: 13 of 15 generated “good” settings produced defect-free parts, while all 5 generated “defective” settings reliably triggered defects, supporting exploration of the process window. The proposed generative–surrogate framework enables rapid adaptation to environmental changes and can be extended to other manufacturing inverse design problems.