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