Energy-Adaptive Diffusion via Dynamic Token Pruning for Carbon-Efficient On-Device Generation

Mahule Roy, Subhas Roy

Keywords: energy-adaptive diffusion, token pruning, adaptive step scheduling, latent diffusion, on-device generative AI, energy-efficient image synthesis

TL;DR: We introduce EAD-Diff, a diffusion model that dynamically prunes latent tokens and adjusts denoising steps to minimize energy while preserving image quality on edge devices.

Abstract: Diffusion-based generative models achieve state-of-the-art image synthesis but are computationally intensive, limiting deployment on mobile and edge devices. Existing efficiency techniques typically rely on fixed inference schedules and static compute budgets, failing to account for dynamic device conditions such as battery level or thermal constraints. We introduce Energy-Adaptive Diffusion (EAD-Diff), a framework that dynamically adapts computation to device-level energy availability. EAD-Diff combines budget-conditioned token pruning in latent feature space with adaptive denoising step scheduling, trained under a unified objective that balances perceptual quality and energy consumption. We evaluate our approach on CIFAR-10 and CelebA-HQ 256$\times$256, measuring energy and latency on NVIDIA Jetson Orin Nano and Raspberry Pi 5. Results show up to 38\% energy reduction with minimal FID degradation, demonstrating that runtime-adaptive diffusion is a practical pathway toward sustainable, on-device generative AI.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 3