Image generation methods, including inpainting, are evolving rapidly; however, high memory requirements continue to limit their practical deployment. As a result, the efficient utilization of Latent Diffusion Models on edge devices has become increasingly important. This work explores techniques for reducing memory usage in Latent Diffusion Models while preserving their generative capabilities.

We propose a resource-efficient inpainting method optimized for CPU-based inference, based on a combination of VAE tiling, attention slicing, and dynamic region-of-interest slicing. Experimental results demonstrate that the model's memory footprint can be significantly reduced while maintaining output quality, without substantial increases in computation time, enabling execution on systems with as little as 4 GB of memory and only two processing cores. While the introduced optimizations, particularly those based on localized image processing, introduce an inherent trade-off between memory usage and computational cost, resulting in longer inference times compared to GPU-accelerated solutions, they demonstrate strong potential for deployment in memory-limited environments.

Additionally, we provide analysis of key deployment bottlenecks, including model compilation for cold-start overhead mitigation, proper runtime configuration and scheduler selection. These findings confirm the feasibility of effectively deploying Latent Diffusion Models for inpainting tasks on CPU-only, resource-constrained platforms, thereby broadening their applicability to edge computing scenarios.