LoRA (Low-Rank Adaptation)

What Are LoRAs?

LoRAs are small model add-ons that modify an AI processor's behavior without replacing the entire model. They can add specific styles, subjects, concepts, or aesthetic qualities to your renders.

Think of LoRAs as "style packs" — the base model does the heavy lifting, and LoRAs fine-tune the output in specific directions.

Supported Processors

LoRAs are available for these processors: - SDXL — Largest LoRA ecosystem, thousands available online - SD 1.5 — Massive community library, longest-established ecosystem - SD 2.1 — Smaller selection, compatible with 2.1-specific LoRAs - Qwen-Image — Emerging ecosystem, newer LoRAs being developed

Note: SD 3.5 and TeleStyle do not currently support external LoRAs.

Using LoRAs

Adding LoRAs to Your Library

1. Open the LoRA tab in configuration
2. Browse the built-in catalog (15 curated LoRAs included)
3. Or download from HuggingFace — paste a model URL
4. Or import local .safetensors files

Applying LoRAs to a Render

1. Select a compatible processor (SDXL, SD 1.5, SD 2.1, or Qwen-Image)
2. Open the LoRA tab
3. Toggle on one or more LoRAs
4. Adjust weight for each (0.0 to 1.0+)
5. 0.5 = subtle influence
6. 0.75 = moderate influence
7. 1.0 = full effect
8. Multiple LoRAs stack — their effects combine

Multi-LoRA Stacking

You can apply multiple LoRAs simultaneously. Their effects combine additively. Tips: - Keep total weights reasonable (sum ≤ 1.5 for stability) - Complementary LoRAs work best (style + detail, not two competing styles) - Order doesn't matter — LoRAs are applied simultaneously

Built-In LoRA Catalog

15 curated LoRAs ship with the application: - 9 SDXL LoRAs (styles, aesthetics, detail enhancement) - 3 SD 1.5 LoRAs (classic styles) - 3 SD 3.5 LoRAs (newer styles)

IP-Adapters

Two IP-Adapter models available for image-guided generation: - SD 1.5 IP-Adapter — Use a reference image to guide style/composition - SDXL IP-Adapter — Same concept for SDXL processor

IP-Adapters let you "show" the AI a reference image instead of describing what you want in text.

Tips

• LoRA weight 0.6-0.8 is usually the sweet spot
• Some LoRAs have trigger words — check the model description
• If results are oversaturated or distorted, reduce weight
• LoRAs trained for one processor won't work with another

See also: Pipelines for processor details | Model Management for downloading LoRAs