Plug-and-Play Diffusion Distillation

TL;DR

This paper introduces a lightweight guide model for diffusion-based image generation that significantly speeds up inference, maintains high image quality, and is adaptable across different models without retraining.

Contribution

The authors propose a novel distillation method that creates a plug-and-play guide for diffusion models, reducing inference time and parameter count while preserving image quality.

Findings

Reduces inference computation by nearly 50%.

Requires only 1% of the original model's trainable parameters.

Achieves comparable FID scores with 8-16 steps.

Abstract

Diffusion models have shown tremendous results in image generation. However, due to the iterative nature of the diffusion process and its reliance on classifier-free guidance, inference times are slow. In this paper, we propose a new distillation approach for guided diffusion models in which an external lightweight guide model is trained while the original text-to-image model remains frozen. We show that our method reduces the inference computation of classifier-free guided latent-space diffusion models by almost half, and only requires 1\% trainable parameters of the base model. Furthermore, once trained, our guide model can be applied to various fine-tuned, domain-specific versions of the base diffusion model without the need for additional training: this "plug-and-play" functionality drastically improves inference computation while maintaining the visual fidelity of generated images. Empirically, we show that our approach is able to produce visually appealing results and achieve a comparable FID score to the teacher with as few as 8 to 16 steps.