Aiming to address the problems of traditional image generation models, including insufficient latent space representation capability in complex scenes and low fidelity in high-resolution image generation, this study proposes a two-stage training framework based on an improved vector quantized variational autoencoder (IVQ-VAE) and a feature-fused Transformer diffusion (FFTD) model. By introducing the attention mechanism, residual blocks, and multi-component loss function, IVQ-VAE significantly enhances the semantic representation capability of the latent space and fidelity of generated images, overcoming the limitations of traditional encoders in capturing complex image features. Built on a Transformer architecture, FFTD further improves the modeling capacity of complex image structures by integrating multi-resolution sampling and adaptive feature fusion. The two-stage training strategy first pre-trains IVQ-VAE to generate high-quality latent representations, then freezes its parameters, and trains FFTD by employing a denoising diffusion implicit model (DDIM) to optimize the noise prediction and image generation process. This framework achieves significant improvements in detail fidelity and visual quality of the generated images on datasets such as CelebA-HQ and AFHQ, validating its effectiveness in high-resolution image generation.