Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2507.02608 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914127210872832 |
|---|---|
| author | Rozet, François Ohana, Ruben McCabe, Michael Louppe, Gilles Lanusse, François Ho, Shirley |
| author_facet | Rozet, François Ohana, Ruben McCabe, Michael Louppe, Gilles Lanusse, François Ho, Shirley |
| contents | The steep computational cost of diffusion models at inference hinders their use as fast physics emulators. In the context of image and video generation, this computational drawback has been addressed by generating in the latent space of an autoencoder instead of the pixel space. In this work, we investigate whether a similar strategy can be effectively applied to the emulation of dynamical systems and at what cost. We find that the accuracy of latent-space emulation is surprisingly robust to a wide range of compression rates (up to 1000x). We also show that diffusion-based emulators are consistently more accurate than non-generative counterparts and compensate for uncertainty in their predictions with greater diversity. Finally, we cover practical design choices, spanning from architectures to optimizers, that we found critical to train latent-space emulators. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2507_02608 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation Rozet, François Ohana, Ruben McCabe, Michael Louppe, Gilles Lanusse, François Ho, Shirley Machine Learning Fluid Dynamics The steep computational cost of diffusion models at inference hinders their use as fast physics emulators. In the context of image and video generation, this computational drawback has been addressed by generating in the latent space of an autoencoder instead of the pixel space. In this work, we investigate whether a similar strategy can be effectively applied to the emulation of dynamical systems and at what cost. We find that the accuracy of latent-space emulation is surprisingly robust to a wide range of compression rates (up to 1000x). We also show that diffusion-based emulators are consistently more accurate than non-generative counterparts and compensate for uncertainty in their predictions with greater diversity. Finally, we cover practical design choices, spanning from architectures to optimizers, that we found critical to train latent-space emulators. |
| title | Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation |
| topic | Machine Learning Fluid Dynamics |
| url | https://arxiv.org/abs/2507.02608 |