Trajectory Stitching for Solving Inverse Problems with Flow-Based Models
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2026
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866912890533969920 |
|---|---|
| author | Denker, Alexander Eliasof, Moshe Kereta, Zeljko Schönlieb, Carola-Bibiane |
| author_facet | Denker, Alexander Eliasof, Moshe Kereta, Zeljko Schönlieb, Carola-Bibiane |
| contents | Flow-based generative models have emerged as powerful priors for solving inverse problems. One option is to directly optimize the initial latent code (noise), such that the flow output solves the inverse problem. However, this requires backpropagating through the entire generative trajectory, incurring high memory costs and numerical instability. We propose MS-Flow, which represents the trajectory as a sequence of intermediate latent states rather than a single initial code. By enforcing the flow dynamics locally and coupling segments through trajectory-matching penalties, MS-Flow alternates between updating intermediate latent states and enforcing consistency with observed data. This reduces memory consumption while improving reconstruction quality. We demonstrate the effectiveness of MS-Flow over existing methods on image recovery and inverse problems, including inpainting, super-resolution, and computed tomography. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_08538 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Trajectory Stitching for Solving Inverse Problems with Flow-Based Models Denker, Alexander Eliasof, Moshe Kereta, Zeljko Schönlieb, Carola-Bibiane Image and Video Processing Machine Learning 68T07 Flow-based generative models have emerged as powerful priors for solving inverse problems. One option is to directly optimize the initial latent code (noise), such that the flow output solves the inverse problem. However, this requires backpropagating through the entire generative trajectory, incurring high memory costs and numerical instability. We propose MS-Flow, which represents the trajectory as a sequence of intermediate latent states rather than a single initial code. By enforcing the flow dynamics locally and coupling segments through trajectory-matching penalties, MS-Flow alternates between updating intermediate latent states and enforcing consistency with observed data. This reduces memory consumption while improving reconstruction quality. We demonstrate the effectiveness of MS-Flow over existing methods on image recovery and inverse problems, including inpainting, super-resolution, and computed tomography. |
| title | Trajectory Stitching for Solving Inverse Problems with Flow-Based Models |
| topic | Image and Video Processing Machine Learning 68T07 |
| url | https://arxiv.org/abs/2602.08538 |