LoRA (Low-Rank Adaptation) is a small “patch” applied to a base model. A LoRA file is usually 50–200 MB and adds a specific character, style or concept.
Basic syntax
a young woman, <lora:name:0.85>, in a cafe, photorealistic
- <lora:filename:weight> is parsed by the WebUI
- 0.6–0.9 is the sweet spot for most LoRAs
- Some LoRAs require a trigger word at the start of the prompt
LoRA types
| Type | Purpose | Recommended weight |
|---|---|---|
| Character LoRA | Fix one specific character | 0.7–0.9 |
| Style LoRA | Apply a specific style | 0.5–0.8 |
| Concept LoRA | Add an object or pose concept | 0.6–0.9 |
| Detail-enhancer LoRA | Boost a specific kind of detail (skin, fabric) | 0.3–0.6 |
Stacking multiple LoRAs
<lora:char-mei:0.85> <lora:style-watercolor:0.5> <lora:detail-skin:0.4> a young female violinist in a cafe
Rules:
- Total weight should stay below 2.0 (1.5 is the recommended cap)
- Avoid two LoRAs targeting the same axis (two realistic-skin LoRAs will fight)
- Character > style > detail in weight priority
Trigger words
Many LoRA authors train with a specific keyword that activates the effect. For example, a Ghibli-style LoRA might need the trigger word "ghibli style" in the prompt to apply. The Civitai page usually lists trigger words.
Common gotchas
Base model mismatch
SDXL LoRAs do not work on SD 1.5 base, and vice versa. Flux is yet another ecosystem.
Weight too high
Above 1.0 the LoRA over-fits and the image gets rigid. 0.6–0.9 is the safe range.
Forgetting the trigger word
LoRA loaded but no effect — the trigger word is usually missing.