Reproduce “Low-Rank Adaptation of Neural Fields” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anh Truong, Ahmed H. Mahmoud, Mina Konaković Luković, Justin Solomon
LoRA represents edits to instance-specific neural fields by freezing the original network and learning a compact, rank-limited weight update. In image experiments, the red-marker edit reaches 41.31 dB PSNR versus 42.12 dB for full fine-tuning, with 46k rather than 340k updated parameters. This supports parameter efficiency for tested minor edits, not faster training or large changes.
Key insights
- Small changes or perturbations in the output function of a neural field correspond to low-rank updates in the network's weight space.
- LoRA can effectively capture these low-rank updates for instance-specific neural fields, extending parameter-efficient fine-tuning beyond large pre-trained foundation models.
- The online, reconstruction-loss-aware factorization of LoRA during training is more effective for encoding neural field updates compared to post-hoc low-rank factorizations, leading to superior visual quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.