AI & Computation

LoRA

Low-Rank Adaptation — a lightweight fine-tuning method that specializes large AI models for a narrow domain.

AI & Computation 1 min read
Definition

LoRA enables training small adapter layers on top of a base model — for example, teaching a renderer the visual signature of Indo-Saracenic architecture without retraining the entire model.

Why it matters

Full fine-tuning of a diffusion model costs thousands of dollars. A LoRA can be trained for under $20 and shipped as a 50–200 MB file, making style customisation practical for studios.

Key points

  • Adapter sits alongside the base model; original weights are untouched.
  • Multiple LoRAs can be combined with weighted blending.
  • Typical training: 30–100 reference images, 1–2 hours on a single GPU.

Examples

  • A LoRA trained on 60 photos of a studio's prior projects to keep their house style consistent.
  • A region-specific LoRA for Kerala vernacular roofs.
In AI Cadbull Studio

Cadbull ships curated style LoRAs (Modern Indian, Indo-Saracenic, Tropical Vernacular) you can mix into any render.

Try it now

Ready to try AI Cadbull Studio?

Paid plans from $9.99 include 50–1500 AI Credits. every AI tool included.