Imagine Flash: Accelerating Emu Diffusion Models with Backward Distillation

TL;DR

This paper introduces Imagine Flash, a novel distillation framework that accelerates diffusion models to generate high-quality, diverse images in just one to three steps, significantly reducing inference time without sacrificing quality.

Contribution

It presents a new backward distillation method with shift-adaptive loss and noise correction, enabling high-fidelity diffusion model distillation in extremely low-step regimes.

Findings

Outperforms existing methods in quantitative metrics and human evaluations.

Achieves performance comparable to the original model with only three denoising steps.

Enables efficient high-quality image generation with minimal inference steps.

Abstract

Diffusion models are a powerful generative framework, but come with expensive inference. Existing acceleration methods often compromise image quality or fail under complex conditioning when operating in an extremely low-step regime. In this work, we propose a novel distillation framework tailored to enable high-fidelity, diverse sample generation using just one to three steps. Our approach comprises three key components: (i) Backward Distillation, which mitigates training-inference discrepancies by calibrating the student on its own backward trajectory; (ii) Shifted Reconstruction Loss that dynamically adapts knowledge transfer based on the current time step; and (iii) Noise Correction, an inference-time technique that enhances sample quality by addressing singularities in noise prediction. Through extensive experiments, we demonstrate that our method outperforms existing competitors in quantitative metrics and human evaluations. Remarkably, it achieves performance comparable to the teacher model using only three denoising steps, enabling efficient high-quality generation.