CVPR 2026 Poster

Weight Space Representation Learning
via Neural Field Adaptation

Zhuoqian Yang, Mathieu Salzmann, Sabine Süsstrunk

IVRL · EPFL

Also to be presented at the ICML 2026 Workshop on Weight-Space Symmetries.

paper code video poster (pdf)
Jul 10

Workshop poster session

Fri, Jul 10, 2026 · 3:30–5:00 PM KST (Seoul). Join the Zoom room to chat with me during the ICML 2026 Workshop on Weight-Space Symmetries poster session.

open zoom

https://epfl.zoom.us/j/62457584373?pwd=Dp9F8ADGEcCGunzxVuNIBWlIxFXBpM.1

The Zoom link will appear here on the day of the session.

§ 01

TL;DR

t-SNE of weight-space representations, colored by object category and sized by perturbation strength across MLP, LoRA, and mLoRA parameterizations
1

Properly constrained weights exhibit semantic structure and serve as effective data representations.

Qualitative comparison of unconditional generation: HyperDiffusion versus ours on ShapeNet shapes and FFHQ faces
2

Structured weight space enables effective weight space generation.

§ 02

Video

§ 03

Poster

CVPR 2026 poster: Weight Space Representation Learning via Neural Field Adaptation
§ 04

Cite this work

@inproceedings{yang2026wsr,
  title     = {Weight Space Representation Learning via Neural Field Adaptation},
  author    = {Yang, Zhuoqian and Salzmann, Mathieu and S{\"u}sstrunk, Sabine},
  booktitle = {Proceedings of the IEEE/CVF Conference on
               Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}