By Lucas Desfossez · Rawrr project operator
A LoRA is a lightweight adaptation used with a compatible base model to influence outputs toward a learned style, subject or task. For creative AI, it can make a workflow more personal without replacing the whole base model.
Using a LoRA is not the same as training one
Using an existing LoRA means loading its weights into a supported workflow. Training a new one means preparing an appropriate dataset and running a separate adaptation process. Those tasks can have different hardware, tooling and licence requirements. A tool that supports inference with LoRAs does not automatically support training them.
Check compatibility first
Confirm the base-model family and version, expected loader and workflow settings. A LoRA made for one model family is not universally interchangeable with another. Start with the author’s documented settings, change one variable at a time and compare results against a baseline without the adaptation.
Build a repeatable creative test
- Choose a reference task and inputs you are allowed to use.
- Record the base model and LoRA versions.
- Keep prompt, seed and output dimensions fixed while varying LoRA strength.
- Compare style fidelity against unwanted changes to shape, composition or identity.
- Save the selected configuration with the output so another test can reproduce it.
Commercial rights still matter
Review the base-model licence, the LoRA licence and your rights to any source material. A file being downloadable does not establish permission for every commercial use. Client-specific training needs permission to use those references, not just possession of the files.
Keep private adaptations private
Local inference can use a compatible LoRA without uploading prompts and source assets to a remote generation service. Training has its own data boundary: check where that process runs. Protect confidential datasets and weights, including backups. Do not upload a client adaptation to a public model library unless explicitly authorized.
LoRAs in Rawrr
Personalized LoRA workflows are part of Rawrr’s creative direction for a local-first platform. The beta will establish supported base models and workflows. This is not a claim that every LoRA, or every local training method, is already supported. Tell us which workflow you want to test when applying.