Saved in:
Bibliographic Details
Main Authors: Chen, Weiye, Zhu, Qingen, Long, Qian
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2504.08451
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908313918111744
author Chen, Weiye
Zhu, Qingen
Long, Qian
author_facet Chen, Weiye
Zhu, Qingen
Long, Qian
contents Recent advances in visual synthesis have leveraged diffusion models and attention mechanisms to achieve high-fidelity artistic style transfer and photorealistic text-to-image generation. However, real-time deployment on edge devices remains challenging due to computational and memory constraints. We propose Muon-AD, a co-designed framework that integrates the Muon optimizer with attention distillation for real-time edge synthesis. By eliminating gradient conflicts through orthogonal parameter updates and dynamic pruning, Muon-AD achieves 3.2 times faster convergence compared to Stable Diffusion-TensorRT, while maintaining synthesis quality (15% lower FID, 4% higher SSIM). Our framework reduces peak memory to 7GB on Jetson Orin and enables 24FPS real-time generation through mixed-precision quantization and curriculum learning. Extensive experiments on COCO-Stuff and ImageNet-Texture demonstrate Muon-AD's Pareto-optimal efficiency-quality trade-offs. Here, we show a 65% reduction in communication overhead during distributed training and real-time 10s/image generation on edge GPUs. These advancements pave the way for democratizing high-quality visual synthesis in resource-constrained environments.
format Preprint
id arxiv_https___arxiv_org_abs_2504_08451
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Muon-Accelerated Attention Distillation for Real-Time Edge Synthesis via Optimized Latent Diffusion
Chen, Weiye
Zhu, Qingen
Long, Qian
Computer Vision and Pattern Recognition
Recent advances in visual synthesis have leveraged diffusion models and attention mechanisms to achieve high-fidelity artistic style transfer and photorealistic text-to-image generation. However, real-time deployment on edge devices remains challenging due to computational and memory constraints. We propose Muon-AD, a co-designed framework that integrates the Muon optimizer with attention distillation for real-time edge synthesis. By eliminating gradient conflicts through orthogonal parameter updates and dynamic pruning, Muon-AD achieves 3.2 times faster convergence compared to Stable Diffusion-TensorRT, while maintaining synthesis quality (15% lower FID, 4% higher SSIM). Our framework reduces peak memory to 7GB on Jetson Orin and enables 24FPS real-time generation through mixed-precision quantization and curriculum learning. Extensive experiments on COCO-Stuff and ImageNet-Texture demonstrate Muon-AD's Pareto-optimal efficiency-quality trade-offs. Here, we show a 65% reduction in communication overhead during distributed training and real-time 10s/image generation on edge GPUs. These advancements pave the way for democratizing high-quality visual synthesis in resource-constrained environments.
title Muon-Accelerated Attention Distillation for Real-Time Edge Synthesis via Optimized Latent Diffusion
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2504.08451