Skip to main content

Research · Technical

LoRA: Low-Rank Adaptation of Large Language Models

ICLR 2022 · Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W. · 2022

Key points

  1. LoRA fine-tunes large models by learning low-rank update matrices for a subset of weight matrices, reducing trainable parameters by orders of magnitude while preserving output quality.
  2. The technique has been adopted as the dominant fine-tuning mechanism in open-source diffusion model ecosystems, underpinning character LoRAs and style LoRAs used across animation pre-production pipelines.
  3. LoRA underpins character and style fine-tuning across open-source diffusion pipelines, making parameter-efficient adaptation the practical mechanism for character consistency in animation pre-production.

Summary

Hu et al. (2021, published ICLR 2022) propose Low-Rank Adaptation, a parameter-efficient fine-tuning method that inserts trainable low-rank matrix pairs into the weight updates of a frozen pretrained model, dramatically reducing the compute and storage cost of task-specific adaptation. Although the original paper targets large language models, LoRA was rapidly adopted for diffusion models and is now the dominant mechanism for character and style fine-tuning in open-source image-generation pipelines. Character LoRAs and style LoRAs are in active use across animation pre-production, making this a foundational text for understanding how generative tools are customised in practice.

Source

Source: ICLR 2022 ↗ (Research)

Cite this item

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L. & Chen, W. (2022). ‘LoRA: Low-Rank Adaptation of Large Language Models’, ICLR 2022. Available at: https://arxiv.org/abs/2106.09685

Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.