A LoRA is a lightweight AI model add-on (under 150 MB) that fine-tunes image generation toward styles, faces, or concepts. Learn how LoRAs work and how to use them on CRAISEE.
Frequently asked questions
What is a LoRA in AI image generation?
A LoRA (Low-Rank Adaptation) is a lightweight adapter file — typically 1–150 MB — that fine-tunes a base AI image model to consistently produce a specific style, face, or concept. Unlike full model checkpoints (2–12 GB), LoRAs modify only a fraction of the model's parameters, making them fast to download and easy to swap.
How does a LoRA work with Stable Diffusion?
A LoRA attaches to a base model like Stable Diffusion XL or Flux.1 Dev and adjusts a small subset of its weights to bias outputs toward a trained concept. It cannot generate images on its own — it must be layered on top of a compatible base model. The result is targeted style or subject control without retraining the entire model.
What is the difference between a LoRA and a full checkpoint model?
A full checkpoint is a complete AI model (2–12 GB) capable of generating images independently. A LoRA is a small add-on file (1–150 MB) that modifies an existing checkpoint to steer outputs toward a specific style or concept. LoRAs are faster to download, easier to share, and can be combined, whereas checkpoints cannot be stacked the same way.
Can you use multiple LoRAs at the same time?
Yes. On CRAISEE Image Studio, you can stack up to four LoRAs simultaneously on a single base model. Each LoRA can be assigned an individual strength weight, letting you blend styles, combine a character face with an artistic aesthetic, or layer multiple concepts into one generation without switching base models.
What can LoRAs be trained on?
LoRAs can be trained on virtually any visual concept: a specific artistic style (e.g., flat illustration, oil painting), a real or fictional face, a product, a character, an object, or an abstract aesthetic. The training dataset typically consists of 10–30 curated images, making LoRA training far more accessible than full model fine-tuning.
Are LoRAs safe to use and where can I find them?
LoRAs from reputable sources are generally safe, but always download from trusted platforms. On CRAISEE, LoRAs are curated and accessible directly in the Image Studio LoRA selector panel, searchable by name, style tag, or base model compatibility. Third-party sources like Civitai also host large LoRA libraries, though quality and safety vary.
How do I use a LoRA on CRAISEE?
In CRAISEE Image Studio, open the LoRA selector panel located alongside the base model picker in the main generation panel. Search by name, style tag, or base model compatibility, then apply your chosen LoRA. You can stack up to four LoRAs and adjust each one's strength weight individually before generating your image.
What file size is a typical LoRA?
Most LoRA files range from 1 MB to 150 MB, with common sizes around 50–144 MB depending on the training method and rank used. This is dramatically smaller than full model checkpoints, which typically weigh 2–12 GB. The compact size is a core advantage, enabling fast downloads, easy sharing, and flexible stacking.