Research map / Image generation / Text-to-image diffusion
T-LoRA: Single Image Diffusion Model Customization Without Overfitting
Vera Soboleva, Aibek Alanov, Andrey Kuznetsov, Konstantin Sobolev
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add timestep-dependent rank masking (Vanilla T-LoRA) to LoRA, rank 64 (ablation) | LoRA · Single-image (rank 64) | Text similarity (TS): 0.232 -> 0.240 | Improved |
| add Ortho-LoRA orthogonal init to Vanilla T-LoRA, rank 64 (ablation) | T-LoRA · Single-image (rank 64) | Text similarity (TS): 0.240 -> 0.256 | Improved |
| add Ortho-LoRA init to Vanilla T-LoRA, rank 4 (ablation) | T-LoRA · Single-image (rank 4) | Text similarity (TS): 0.259 -> 0.255 | No clear effect |
| T-LoRA vs LoRA text similarity, rank 64 | LoRA · Single-image (rank 64) | Text similarity (TS): 0.232 -> 0.256 | Improved |
| T-LoRA vs LoRA image similarity, rank 64 | LoRA · Single-image (rank 64) | Image similarity (IS): 0.901 -> 0.900 | No clear effect |
| fine-tune only at noisiest timesteps [800,1000] (ablation) | T-LoRA · SD-XL LoRA | qualitative overfitting | Hurt |
| exclude noisiest timesteps from fine-tuning (ablation) | T-LoRA · SD-XL LoRA | concept fidelity | Hurt |
| lower rmin from 50% to 25% of rank (ablation) | T-LoRA · T-LoRA rank 64 | concept fidelity | Mixed |
| initialize with last SVD components of random matrix instead of principal components of W (ablation) | T-LoRA · Ortho-LoRA | TS / IS | Improved |
| T-LoRA vs LoRA with 1, 2, 3 images | LoRA · Multi-image customization | Text similarity (TS): 0.232 -> 0.256 (1 image) | Improved |
| T-LoRA vs LoRA in human concept preservation | LoRA · User study | Concept preservation (%): 39.3 vs 60.7 | Hurt |
| T-LoRA vs SVDiff on text similarity | SVDiff · Single-image customization | Text similarity (TS): 0.256 vs 0.295 | Hurt |
Closest papers by meaning
- Orthogonal Adaptation for Modular Customization of Diffusion Models (CVPR 2024 · similarity 0.92)
- LoRACLR: Contrastive Adaptation for Customization of Diffusion Models (CVPR 2025 · similarity 0.91)
- TimeStep Master: Asymmetrical Mixture of Timestep LoRA Experts for Versatile and Efficient Diffusion Models in Vision (ICML 2025 · similarity 0.91)
- Time-Varying LoRA: Towards Effective Cross-Domain Fine-Tuning of Diffusion Models (NeurIPS 2024 · similarity 0.91)
- PaRa: Personalizing Text-to-Image Diffusion via Parameter Rank Reduction (ICLR 2025 · similarity 0.90)
- Navigating Text-To-Image Customization: From LyCORIS Fine-Tuning to Model Evaluation (ICLR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
