Samsung R&D Institute UK (SRUK) is thrilled to announce that its researchers have won the Best Paper award at the ICCV 2025 Personalization in Generative AI Workshop for their paper titled “LoRA.rar: Learning to Merge LoRAs via Hypernetworks for Subject-Style Conditioned Image Generation”. The proposed LoRA.rar method not only improves image quality, but also achieves a remarkable speedup of over 4000× in the LoRA merging process via its innovative use of parameter generating networks, called hypernetworks.

LoRA.rar overview. LoRA adapters representing a user’s personal object (e.g. my dog) and personal style (my style) are provided and merged via a hypernetwork. The merged LoRA adapter can then generate images of the user’s dog in the user’s style with the selected actions (e.g. jumping and playing cards).

The work was done in collaboration with researchers Ondrej Bohdal and Mete Ozay from the SRUK AI Research Team. The paper was also presented in the main track of the ICCV 2025 conference, the premier conference in the field of computer vision, and published in the following Samsung Research blog:
https://research.samsung.com/blog/LoRA-rar-Learning-to-Merge-LoRAs-via-Hypernetworks-for-Subject-Style-Conditioned-Image-Generation.

Best paper award certificate, for work done in collaboration with SRUK researchers Ondrej Bohdal and Mete Ozay.