Transformed Low-rank Adaptation via Tensor Decomposition and Its Applications to Text-to-image Models

TL;DR

This paper introduces a novel PEFT method combining transform and residual adaptations with tensor decompositions, significantly improving fine-tuning efficiency and performance in text-to-image models like Stable Diffusion.

Contribution

It proposes a new PEFT approach that reduces approximation gaps in LoRA by applying a full-rank transform and residual adaptation with tensor decompositions, enhancing efficiency and accuracy.

Findings

Outperforms LoRA and baselines in fine-tuning Stable Diffusion.

Achieves higher parameter efficiency with better performance.

Reduces approximation error in model adaptation.

Abstract

Parameter-Efficient Fine-Tuning (PEFT) of text-to-image models has become an increasingly popular technique with many applications. Among the various PEFT methods, Low-Rank Adaptation (LoRA) and its variants have gained significant attention due to their effectiveness, enabling users to fine-tune models with limited computational resources. However, the approximation gap between the low-rank assumption and desired fine-tuning weights prevents the simultaneous acquisition of ultra-parameter-efficiency and better performance. To reduce this gap and further improve the power of LoRA, we propose a new PEFT method that combines two classes of adaptations, namely, transform and residual adaptations. In specific, we first apply a full-rank and dense transform to the pre-trained weight. This learnable transform is expected to align the pre-trained weight as closely as possible to the desired weight, thereby reducing the rank of the residual weight. Then, the residual part can be effectively approximated by more compact and parameter-efficient structures, with a smaller approximation error. To achieve ultra-parameter-efficiency in practice, we design highly flexible and effective tensor decompositions for both the transform and residual adaptations. Additionally, popular PEFT methods such as DoRA can be summarized under this transform plus residual adaptation scheme. Experiments are conducted on fine-tuning Stable Diffusion models in subject-driven and controllable generation. The results manifest that our method can achieve better performances and parameter efficiency compared to LoRA and several baselines.