Research map / Architectures and efficiency / Attention and transformers
Denoising Vision Transformers
Jiawei Yang, Katie Z Luo, Jiefeng Li, Congyue Deng, Leonidas Guibas, Dilip Krishnan, Kilian Weinberger, Yonglong Tian, Yue Wang
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 DVT denoiser to pre-trained ViTs | DINOv2 ViT-B · VOC2012 segmentation | mIoU: 83.60 -> 84.84 | Improved |
| add DVT to DINOv2-reg | DINOv2-reg · VOC2012 and ADE20k segmentation | mIoU: +0.86 in VOC2012 and +1.12 in ADE20k | Improved |
| add DVT denoiser | DeiT-III, DINO, EVA02, CLIP, DINOv2 · NYUv2 depth | delta_1: DeiT-III 72.16 -> 71.36; DINO abs rel 0.1701 -> 0.1731 | Mixed |
| add DVT denoiser | DINOv2-reg · VOC2007 object discovery (LOST) | corloc: 38.0 -> 56.1 | Improved |
| add DVT denoiser | six ViTs · VOC detection | mAP: DeiT-III 75.8 -> 77.0 | Improved |
| add artifact field G and residual term to the semantics field (ablation) | per-image denoising · VOC12 val KNN segmentation | mIoU: 67.81 -> 70.94 | Improved |
| add residual term on top of F + G (ablation) | per-image denoising · VOC12 val KNN segmentation | mIoU: 70.82 -> 70.94 | No clear effect |
| use transformer block denoiser instead of conv1x1 (ablation) | generalizable denoiser · VOC2012 val | mIoU: 82.15 -> 84.84 | Improved |
| add learnable positional embeddings to the denoiser (ablation) | single Transformer block denoiser · VOC2012 val | mIoU: 84.81 -> 84.84 | No clear effect |
| use more views in per-image neural field denoising (ablation) | DVT stage one · dense prediction tasks | task performance | Improved |
| train denoiser on aggregated-view features only (F-only) instead of full decomposition (ablation) | DVT · dense prediction tasks | task performance | Hurt |
| train denoiser with only 100 to 1000 denoised samples (ablation) | DVT denoiser · dense prediction tasks | task performance | Improved |
Closest papers by meaning
- Improving Interpretation Faithfulness for Vision Transformers (ICML 2024 · similarity 0.90)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.89)
- Vision Transformers with Self-Distilled Registers (NeurIPS 2025 · similarity 0.88)
- Interpreting vision transformers via residual replacement model (NeurIPS 2025 · similarity 0.88)
- Learning Where to Edit Vision Transformers (NeurIPS 2024 · similarity 0.87)
- Alias-Free ViT: Fractional Shift Invariance via Linear Attention (NeurIPS 2025 · similarity 0.87)
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.
