In complex indoor environments, ultra-wideband multiple-input multiple-output (UWB-MIMO) systems often suffer from limited channel estimation accuracy, which decreases communication capacity and sensing fidelity. To address this challenge and enhance joint channel estimation and capacity, a deep generative model based on a variational autoencoder was developed in this study. The model processes complex-valued channel matrices by separating real and imaginary components, and employs an encoder–decoder structure to learn the latent probabilistic distribution of multipath signatures. Simulation results showed that the variational autoencoder (VAE)-based model outperforms conventional least squares, minimum mean square error, and compressive sensing-based models across the entire signal-to-noise (SNR) ratio range. The model showed substantial gains in low-SNR regions, where sensing fidelity became vulnerable to noise, enhancing channel capacity to 10% at a typical SNR of 10 dB. By accurately recovering spatial structures while suppressing noise, the framework preserves critical multipath features essential for centimeter-level localization and precise ranging. The developed method utilizes UWB-MIMO to enable ultra-reliable communication and high-resolution sensing. These capabilities make VAE-based models applicable to Industrial IoT and smart manufacturing environments, where reliable connectivity and fine-grained environmental awareness are essential for asset tracking, worker safety monitoring, and material characterization.