04. Style with LoRA & Reference

What LoRA adds

LoRA (Low-Rank Adaptation) injects a small weight delta into the base model. Style LoRAs affect palette, line quality, and texture; character LoRAs anchor identity. Weights typically range 0.4–0.9 — start at 0.6 and adjust.

Workflow

1. Pick a base checkpoint (SDXL or SD 3.5).
2. Add one style LoRA at ~0.6.
3. Include trigger tokens from the LoRA card in the prompt.
4. Save a reusable template: {trigger}, {subject}, {lighting}, {camera}.

Combining LoRA with references

Use img2img or ControlNet when you need layout fidelity plus LoRA styling. Keep img2img strength lower (0.35–0.55) so the LoRA defines look while the reference holds composition.

Avoid stacking many LoRAs at high weight — interactions become unpredictable.

Building a series

Consistent series share: same LoRA weight, similar lighting vocabulary, fixed aspect ratio, and optionally fixed seed per character. Vary only the subject clause between images.

Practice

Create a 4-image series with the same LoRA and different subjects (fruit, vehicle, portrait, interior). Keep weight and lighting fixed; only change the subject block.