LoRA

A parameter-efficient fine-tuning technique, introduced by Hu et al. in 2022, in which a large pretrained neural network is specialised by training only a pair of small low-rank matrices that modify specific weight projections, while the original weights remain frozen. LoRA typically trains well under 1% of the original parameters, producing compact adapter files that can be shared and swapped at inference time. In accessibility contexts, LoRA makes it feasible to steer large vision or language models toward specialised styles (such as Easy Read pictograms, sign-language illustrations, or simplified reading) without the compute cost of full retraining, and to stack multiple adapters for different user populations or contexts.