Research map / Label-efficient and robust learning / Test-time adaptation
TTT-MIM: Test-Time Training with Masked Image Modeling for Denoising Distribution Shifts
Youssef Mansour, Xuyang Zhong, Serdar Caglar, Reinhard Heckel
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 |
|---|---|---|---|
| adapt the main head at test time with the pseudo-denoising loss instead of keeping it fixed (ablation) | TTT-MIM · Average over tasks (batch adaptation) | PSNR: 34.36 -> 36.52 | Improved |
| adapt the whole model rather than only group-norm affine parameters (ablation) | TTT-MIM · Average over tasks (batch adaptation) | PSNR: 33.89 -> 36.52 | Improved |
| use pseudo-denoising loss instead of Noise2Self loss for the main head (ablation) | TTT-MIM · Average over tasks (batch adaptation) | PSNR: 35.58 -> 36.52 | Improved |
| adapt a pretrained network at test time instead of zero-shot training from scratch (vs Self2Self) | Single-image denoising · Average over tasks (excluding DND) | PSNR: 33.14 -> 34.26 | Improved |
| adapt a pretrained network with 8 iterations instead of Self2Self training for 150k iterations | Single-image denoising · Computational cost | time per image: 1.2 hours -> less than a second | Improved |
| apply single-image TTT-MIM to Salt & Pepper and Poisson noise shifts | TTT-MIM · ImageNet Poisson noise | PSNR: 32.66 -> 32.91 | No clear effect |
| adapt to a batch of images instead of a single image | TTT-MIM · SIDD | PSNR: 33.58 -> 39.69 | Improved |
| choose the optimal early-stopping iteration per dataset using ground truth instead of fixed 8 | TTT-MIM · Table 1 settings | PSNR: +0.5 dB average | Improved |
| TTT-MIM vs zero-shot denoisers on CT and fastMRI | TTT-MIM · fastMRI and CT (single image) | PSNR: fastMRI 29.87 vs 31.99 (Self2Self) | Hurt |
Closest papers by meaning
- Masked Pre-training Enables Universal Zero-shot Denoiser (NeurIPS 2024 · similarity 0.89)
- Robust Test-Time Adaptation for Single Image Denoising Using Deep Gaussian Prior (ICCV 2025 · similarity 0.89)
- IDR: Self-Supervised Image Denoising via Iterative Data Refinement (CVPR 2022 · similarity 0.89)
- Masked Image Training for Generalizable Deep Image Denoising (CVPR 2023 · similarity 0.88)
- Zero-Shot Noise2Mean: Gap Minimization for Efficient Denoising from a Single Noisy Image (AAAI 2025 · similarity 0.88)
- Zero-shot Denoising via Neural Compression: Theoretical and algorithmic framework (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.
