LoRAs and Fine-Tuning, Explained

The elevator pitch

A LoRA is a small file (usually 50–200 MB) that patches a base image model to specialize in one thing — a specific person's face, a specific art style, a specific character. You load it on top of the base model at generation time.

A full fine-tune modifies the base model itself. Bigger file (2–7 GB), harder to train, harder to combine.

When to use a LoRA

• You want a specific character to appear consistently
• You want an artist's style you can't get from prompting alone
• You need a specific object (a real product, a real logo) rendered accurately
• You want to layer 2–3 concepts together

When to skip LoRAs entirely

• You can describe what you want in words → use prompting first, it's free
• You want something the base model already does well
• You don't have 30–100 clean reference images

Training basics

Modern SDXL LoRA training needs:

• 20–100 reference images at 1024×1024 (more = better up to ~100)
• Consistent style / subject across images
• Captions (short, factual descriptions) for each
• A GPU with 16+ GB VRAM, or a service that rents one

Training takes 1–4 hours on a 4090. Cost: usually $2–10 rented.

The stacking trick

You can load multiple LoRAs at once, each with its own weight (0.0–1.0). A common recipe: character LoRA at 0.8 + style LoRA at 0.6. Play with weights until it looks right.

Imagoat's approach

Imagoat's AI Avatars starts this pipeline with identity reference packages and monthly creator content planning, with the data model ready for future LoRA training. No .safetensors files, no scripts.