T-LoRA: Single Image Diffusion Model Customization Without Overfitting

Vera Soboleva1,2, Aibek Alanov1,2, Andrey Kuznetsov1, Konstantin Sobolev1,3*
FusionBrain Lab1, HSE University2, MSU3
AAAI 2026
*Indicates Corresponding Author
Paper Code
Method Teaser

We introduce T-LoRA, a Timestep-Dependent Low-Rank Adaptation framework specifically designed for diffusion model personalization. T-LoRA reduces overfitting related to position and background, enabling versatile and enriched generation using only a single object image.

Method Overview

Method Overview

Comparison of training methods: LoRA, the proposed Vanilla T-LoRA, and T-LoRA schemes.

  • T-LoRA tackles overfitting problem related to position and background in few-shot diffusion model personalization, enabling versatile and enriched generation;
  • The method is based on the observation that higher (noisier) diffusion timesteps are more vulnerable to overfitting than lower ones;
  • Key components:
    • Timestep-dependent LoRA rank masking to restrict concept information injection;
    • Orthogonal initialization for efficient exclusion of LoRA components.

FLUX.1 [dev] Results

Flux comparison with LoRA

CLIP IS and TS for LoRA, Vanilla T-LoRA, and T-LoRA across dfferent ranks.

SD-XL Results

Qualitative Comparison

Generation examples for T-LoRA alongside other diffusion model customization baselines.

User Study

User study: Pairwise comparison with T-LoRA.

baselines Comparison

Comparison with other baseline methods.

Comparison with LoRA

CLIP IS and TS for LoRA, Vanilla T-LoRA, and T-LoRA across dfferent ranks.

Video Presentation

BibTeX

@article{soboleva2025t,
  title={T-lora: Single image diffusion model customization without overfitting},
  author={Soboleva, Vera and Alanov, Aibek and Kuznetsov, Andrey and Sobolev, Konstantin},
  journal={arXiv preprint arXiv:2507.05964},
  year={2025}
}