Generative Deep Learning — VAE, GAN or Diffusion Model Build

I will build a generative deep learning model for your data synthesis use case — implementing a Variational Autoencoder (VAE) for structured latent space generation, a GAN for high-fidelity sample generation, or a diffusion model for your image or structured data generation requirement — with training stability controls and quality evaluation metrics. Generative models are among the most technically demanding deep learning architectures to train stably — GANs in particular are notorious for mode collapse and training instability that requires architecture choices, training tricks, and monitoring discipline that most tutorials omit, producing implementations that work on toy datasets and fail on real data.

The build covers architecture implementation appropriate to your data type and generation quality requirements, training stability controls (gradient penalty for GANs, KL annealing for VAEs), generation quality evaluation (FID for images, MMD for tabular), sample visualisation, and a generation function producing new samples on demand.