MEPA: Multi-scale Alignment for VAR with Mixture of Experts

Image generation using autoregressive models has advanced significantly in recent years, but still faces fundamental challenges in learning multi-scale representations. Architectures like Visual AutoRegressive (VAR) introduced a coarse-to-fine generation approach, where lower scales capture global semantics and higher scales model fine details. However, using a shared architecture for all scales creates optimization conflicts: the network must balance contradictory objectives between global abstractions and local precision. Furthermore, the causal process causes early semantic errors to propagate and degrade final quality. To address these issues, MEPA (Multi-scale Expert with Progressive Alignment) emerges, a proposal that integrates a scale-aware mixture of experts (MoE), allowing each level to select the most suitable submodels for its specific task. This decouples representations and reduces interference between scales. Complementarily, external self-supervised learning features are incorporated in the early scales, but not through naive alignment, but with a residual aggregation scheme designed for the VAR paradigm. Experiments on ImageNet 256x256 show that MEPA achieves improved FID compared to the dense baseline, requiring only half the training epochs and a smaller parameter budget, with a marginal increase in computational cost. This advance is not only relevant for academic research but also has practical implications for the development of artificial intelligence for businesses that need to generate high-quality visual content with computational efficiency. For example, in custom computer vision applications, such as product prototype generation or synthetic data augmentation, techniques like MEPA allow optimizing resources without sacrificing realism. At Q2BSTUDIO, we understand that adopting advanced models must be accompanied by a robust infrastructure. That is why we offer AWS and Azure cloud services to deploy these models at scale, and business intelligence services like Power BI to analyze performance. Additionally, we combine comprehensive cybersecurity with AI agents that monitor and protect data in real time. Our team develops custom software that integrates these capabilities, from implementing MoE architectures to automating training processes. The evolution of models like VAR and MEPA demonstrates that artificial intelligence research is maturing towards more efficient and adaptable solutions. In an environment where every resource counts, having a technology partner that offers custom applications and robust cloud services makes the difference. At Q2BSTUDIO, we transform these concepts into real projects, helping companies capitalize on generative AI with a practical and scalable approach.