Seismic porosity inversion is inherently ill-posed and highly sensitive to seismic resolution, geological complexity, and well density, particularly in thin-bedded heterogeneous reservoirs. Although deep-learning-based inversion has shown promising results, most approaches remain deterministic and lack reliable uncertainty quantification. In this study, we examine the combined effects of input representation, Bayesian inference, well density, and seismic resolution on porosity inversion using Bayesian deep learning. Two encoder–decoder networks are developed: WaveNet (waveform-based) and SpecNet (time–frequency spectral-based). Their Bayesian versions, BayeWaveNet and BayeSpecNet, incorporate Monte Carlo dropout for uncertainty estimation. A series of controlled experiments is conducted on a synthetic fluvial–deltaic reservoir model under varying seismic frequencies (20–60 Hz) and well densities, and field seismic data are further used to validate the conclusions. Results show that spectral inputs outperform waveform inputs, especially for thin-bedded channels and deltaic deposits. Bayesian models yield more stable, geologically consistent predictions with reduced noise and meaningful uncertainty estimates. Increasing well density and seismic frequency significantly improves accuracy and reduces uncertainty. Overall, BayeSpecNet achieves the best performance, demonstrating the value of integrating spectral representations and Bayesian learning for reliable, risk-aware reservoir characterization.