Dynamic Mixture-of-Experts for Visual Autoregressive Model

2025-10-08

AI Summary

Visual autoregressive models (VARs) suffer from significant computational redundancy in multi-scale image generation due to repeated invocation of full-parameter Transformers across scales. Method: We propose a dynamic Mixture-of-Experts (MoE) routing mechanism with a scale-aware thresholding strategy, which adaptively activates expert subnetworks based on token-level semantic complexity and current spatial resolution—requiring no additional training and enabling fine-grained computation allocation. Contribution/Results: Our approach is the first to jointly model dynamic sparsification and multi-scale generation while preserving the standard Transformer backbone. It substantially alleviates redundancy during high-resolution stages. Experiments on standard image generation benchmarks show a 20% reduction in FLOPs and an 11% speedup in inference latency, while maintaining image quality—measured by FID and LPIPS—comparable to dense baseline models. This achieves an effective trade-off between computational efficiency and generative fidelity.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Computer Vision: Diffusion Models for Vision
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Visual Autoregressive Models (VAR) offer efficient and high-quality image generation but suffer from computational redundancy due to repeated Transformer calls at increasing resolutions. We introduce a dynamic Mixture-of-Experts router integrated into VAR. The new architecture allows to trade compute for quality through scale-aware thresholding. This thresholding strategy balances expert selection based on token complexity and resolution, without requiring additional training. As a result, we achieve 20% fewer FLOPs, 11% faster inference and match the image quality achieved by the dense baseline.

Problem

• Reduces computational redundancy in Visual Autoregressive Models
• Introduces dynamic Mixture-of-Experts for efficient image generation
• Balances expert selection through scale-aware thresholding strategy

Innovation

• Dynamic Mixture-of-Experts router integrated into VAR
• Scale-aware thresholding balances expert selection automatically
• Reduces FLOPs by 20% while maintaining image quality