LegoDiffusion: Micro-Serving Text-to-Image Diffusion Workflows

2026-04-09

AI Summary

Existing text-to-image diffusion workflows are often treated as black boxes, leading to non-shareable models, coarse-grained resource scheduling, and opaque internal data flows. This work proposes a microservice-based diffusion workflow architecture that decouples the pipeline into loosely coupled model execution nodes, enabling for the first time model-level elastic scaling, cross-request model sharing, and adaptive model parallelism. A distributed scheduling system with explicit model inference management supports dynamic orchestration, independent scaling, and fine-grained resource sharing. Experimental results demonstrate that the proposed approach achieves up to threefold higher request throughput compared to existing systems and can handle burst traffic spikes as high as eight times the baseline load.

Abstract

Text-to-image generation executes a diffusion workflow comprising multiple models centered on a base diffusion model. Existing serving systems treat each workflow as an opaque monolith, provisioning, placing, and scaling all constituent models together, which obscures internal dataflow, prevents model sharing, and enforces coarse-grained resource management. In this paper, we make a case for micro-serving diffusion workflows with LegoDiffusion, a system that decomposes a workflow into loosely coupled model-execution nodes that can be independently managed and scheduled. By explicitly managing individual model inference, LegoDiffusion unlocks cluster-scale optimizations, including per-model scaling, model sharing, and adaptive model parallelism. Collectively, LegoDiffusion outperforms existing diffusion workflow serving systems, sustaining up to 3x higher request rates and tolerating up to 8x higher burst traffic.

Problem

• text-to-image generation
• diffusion workflow
• model serving
• resource management
• model sharing

Innovation

• micro-serving
• diffusion workflow
• model decomposition
• adaptive model parallelism
• model sharing