Decouple-Then-Merge: Finetune Diffusion Models as Multi-Task Learning

TL;DR

This paper introduces a Decouple-then-Merge framework for finetuning diffusion models, where separate models are trained for different timesteps and then merged, leading to improved image generation quality.

Contribution

The work proposes a novel finetuning approach that decouples timestep-specific models and merges them, enhancing diffusion model performance without sacrificing inference efficiency.

Findings

Significant quality improvements on 6 benchmarks including COCO30K, ImageNet1K, and LSUN datasets.

Effective techniques for knowledge sharing during finetuning reduce training interference.

Merged models maintain efficiency while achieving higher generation fidelity.

Abstract

Diffusion models are trained by learning a sequence of models that reverse each step of noise corruption. Typically, the model parameters are fully shared across multiple timesteps to enhance training efficiency. However, since the denoising tasks differ at each timestep, the gradients computed at different timesteps may conflict, potentially degrading the overall performance of image generation. To solve this issue, this work proposes a \textbf{De}couple-then-\textbf{Me}rge (\textbf{DeMe}) framework, which begins with a pretrained model and finetunes separate models tailored to specific timesteps. We introduce several improved techniques during the finetuning stage to promote effective knowledge sharing while minimizing training interference across timesteps. Finally, after finetuning, these separate models can be merged into a single model in the parameter space, ensuring efficient and practical inference. Experimental results show significant generation quality improvements upon 6 benchmarks including Stable Diffusion on COCO30K, ImageNet1K, PartiPrompts, and DDPM on LSUN Church, LSUN Bedroom, and CIFAR10. Code is available at \href{https://github.com/MqLeet/DeMe}{GitHub}.