Noise-Aligned Diffusion Bridge (NADB) fixes endpoint underfitting in CVPR 2026
Resolves noise mismatch that causes drift as diffusion bridges approach target distribution.
A new paper accepted at CVPR 2026 introduces Noise-Aligned Diffusion Bridge (NADB) to solve a critical flaw in diffusion bridge models. These models connect two data distributions (e.g., corrupted to clean images) by mimicking standard diffusion score matching. The authors discovered that this approach causes significant underfitting near the target endpoint (t→0), where the predicted variance and direction drift wildly. The root cause: a large discrepancy in noise levels between the network’s input and its regression target.
NADB tackles this by reformulating the diffusion bridge in two steps. First, a mean network provides a cleaner conditional target. Second, a novel noise-aligned mapping ensures consistent noise levels throughout the process. This eliminates the mismatch and corrects the underfitting. The team validated NADB on multiple image restoration and translation benchmarks, showing clear improvements. The code is publicly available, and the work was accepted at CVPR 2026.
- Identifies endpoint underfitting in diffusion bridges due to noise level mismatch (t→0).
- NADB employs a mean network + noise-aligned mapping to resolve the drift.
- Validated on image restoration and translation tasks; accepted at CVPR 2026.
Why It Matters
Improves diffusion bridge reliability for practical image restoration and translation applications.