Research map / Architectures and efficiency / Pruning and efficient tuning
Transformed Low-rank Adaptation via Tensor Decomposition and Its Applications to Text-to-image Models
Zerui Tao, Yuhta Takida, Naoki Murata, Qibin Zhao, Yuki Mitsufuji
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 TRM transform on top of LoRA rank 4 (ablation) | LoRA · Segmentation-to-image (ADE20K) | mIoU: 27.72 -> 29.23 | Improved |
| add TRM transform on top of LoRA rank 4 (ablation) | LoRA · Segmentation-to-image (ADE20K) | aACC: 64.99 -> 69.21 | Improved |
| add TRM transform on top of LoRA rank 4 (ablation) | LoRA · Canny-to-image (ADE20K) | IoU: 0.180 -> 0.203 | Improved |
| add TRM transform on top of LoRA rank 4 (ablation) | LoRA · Landmark-to-image (CelebA-HQ) | error: 5.32 -> 5.32 | No clear effect |
| add diagonal magnitude (DoRA) to LoRA | LoRA · Landmark-to-image (CelebA-HQ) | error: 5.32 -> 6.43 | Hurt |
| use TR residual rank 8 with TRM instead of LoRA residual, fewer parameters | TLoRA · Segmentation-to-image (ADE20K) | mIoU: 29.23 -> 27.42 | Hurt |
| replace LoRETTA (non-zero init TR) with TLoRA TR residual with zero init | LoRETTA · Landmark-to-image (CelebA-HQ) | error: 118.14 -> 5.84 | Improved |
| add identity regularization on the transform | TLoRA · controllable generation | quality | Improved |
| use TRM transform instead of OFT block-diagonal transform with LoRA (ablation) | LoRA · simulation SDXL weights | approximation error | Improved |
| use TLoRA(2,6) instead of OFT r=32 to cut parameters | OFT · Canny-to-image (ADE20K) | IoU: 0.165 -> 0.184 | Improved |
| use TLoRA with TRM + TR rank 2 for SDXL subject-driven generation instead of LoRA r=1 | LoRA · DreamBooth subject-driven | parameters (M): 1.45 -> 0.40 | Improved |
| add TRM transform to LoRA (ablation) | LoRA · training speed | Iter./Sec.: 0.58 -> 0.37 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ETHER+ — ETHER: Efficient Finetuning of Large-Scale Models with Hyperplane Reflections (ICML 2024) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Decoupling Angles and Strength in Low-rank Adaptation (ICLR 2025 · similarity 0.91)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.91)
- LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation (ICLR 2025 · similarity 0.90)
- The Expressive Power of Low-Rank Adaptation (ICLR 2024 · similarity 0.90)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.90)
- FouRA: Fourier Low-Rank Adaptation (NeurIPS 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.
