UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning
Piotr Wójcik, Maksym Petrenko, Wojciech Gromski, Przemysław Spurek, Maciej Zieba · 2026
Read UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning by Piotr Wójcik, Maksym Petrenko, Wojciech Gromski, Przemysław Spurek, Maciej Zieba (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-03410.txt.
Description
Erasing a concept from an image generator sounds simple, but related ideas often slip through. Piotr Wójcik and 4 co-authors present UnHype in a 2026 paper, a framework for machine unlearning in text-to-image models. It builds on LoRA, a light fine-tuning method, whose fixed adapters adapt poorly to different concept meanings.
UnHype adds a hypernetwork, a small network that generates the LoRA weights on the fly. At run time it reads the CLIP embedding of the request, so the erasure adapts to context. The framework plugs into Stable Diffusion and newer flow-based models. Tests cover removing explicit content, erasing celebrities, and erasing objects, including several concepts erased at the same time.
Catalog record
| CREATOR | Piotr Wójcik, Maksym Petrenko, Wojciech Gromski, Przemysław Spurek, Maciej Zieba |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-03410