Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces LUD-DIF, a novel diffusion-based method for addressing inverse problems using unpaired data, which is often more accessible than paired data. By decoupling the evidence lower bound into two independent diffusion processes, the authors provide a theoretical framework grounded in variational inference, alongside a derived loss function and error analysis. Experimental results show that LUD-DIF significantly outperforms existing methods across various image inverse problems, highlighting its effectiveness and generalization in scenarios with unpaired data.
Unpaired data can be leveraged to achieve state-of-the-art performance in solving complex inverse problems, challenging the reliance on paired datasets.
Data is important in many deep learning-based inverse problem solvers. However, obtaining sufficient paired data in many scenarios remains highly challenging, while unpaired data is cheap. To maximize data utilization, this paper proposes LUD-DIF, a diffusion-based approach for solving inverse problems with unpaired data. Starting from the evidence lower bound (ELBO) of the joint distribution, we decouple it into two independent diffusion processes under the weak-coupling assumption. The method provides theoretical support from a variational inference perspective, derives the loss function, quantitatively analyzes the error bound introduced by the assumption, and offers a theorem-motivated heuristic for hyperparameter selection. Experimental results demonstrate that LUD-DIF achieves outstanding performance on multiple image inverse problems, validating its effectiveness and generalization capability in unpaired inverse problem settings.
Tsinghua AI