Search papers, labs, and topics across Lattice.
The paper introduces Constrained Particle Seeking (CPS), a gradient-free approach for solving diffusion inverse problems without requiring explicit knowledge of the forward observation process. CPS reformulates the inverse problem as a constrained optimization task, actively searching for optimal particles while enforcing constraints that align with high-density regions of the unconditional prior. Experiments on image and scientific inverse problems demonstrate that CPS achieves performance comparable to gradient-based methods and outperforms existing gradient-free approaches.
Ditch the gradients: a new particle-based method solves diffusion inverse problems using only forward passes, matching gradient-based performance.
Diffusion models have gained prominence as powerful generative tools for solving inverse problems due to their ability to model complex data distributions. However, existing methods typically rely on complete knowledge of the forward observation process to compute gradients for guided sampling, limiting their applicability in scenarios where such information is unavailable. In this work, we introduce \textbf{\emph{Constrained Particle Seeking (CPS)}}, a novel gradient-free approach that leverages all candidate particle information to actively search for the optimal particle while incorporating constraints aligned with high-density regions of the unconditional prior. Unlike previous methods that passively select promising candidates, CPS reformulates the inverse problem as a constrained optimization task, enabling more flexible and efficient particle seeking. We demonstrate that CPS can effectively solve both image and scientific inverse problems, achieving results comparable to gradient-based methods while significantly outperforming gradient-free alternatives. Code is available at https://github.com/deng-ai-lab/CPS.
Tsinghua AI