LoRA Image Generator

A LoRA, short for Low-Rank Adaptation, is a small file that nudges a base image model toward one thing: a painting style, a lighting look, a character, an object, or a kind of composition. It is a few dozen to a few hundred megabytes instead of the many gigabytes of a full model, which is why the community trains and shares thousands of them on sites such as CivitAI.

A LoRA image generator runs those files for you. Normally you would download a base model and every LoRA you want, install a local interface, and own a graphics card with enough memory to hold it all. On Mage you pick a base model, add LoRAs from the catalog or import them by link, set each one's weight, and generate in the browser on any device.

LoRAs belong to the model they were trained on, so the base model comes first. On Mage the LoRA-ready image models are MOSS-Image, Stable Diffusion XL and its checkpoint families, Stable Diffusion 1.5, Anima, Z-Image, Krea 2 Turbo and Chroma, and the video models Wan Video 2.2 and MiniMax-H3 take LoRAs too.

Start with one LoRA at a time. Each one pulls the model toward its own training data, and two or three stacked at full weight tend to fight each other and blur into something neither intended. Once one LoRA looks right, add a second at a lower weight, around 0.5, and raise it slowly.

Read the LoRA's page before you use it. Its author usually lists trigger words, a recommended weight, and the base model it was trained on, and shows the prompts behind its sample images. Copying one of those prompts is the quickest way to see what the LoRA does before you write your own.

Keep the rest of the prompt plain. A style LoRA already carries the look, so describe the subject, the setting and the light, and leave the style words to the LoRA. If the result drifts away from the LoRA's look, the prompt is probably naming a competing style.

• One LoRA first, then add more at lower weights.
• Copy the trigger words from the LoRA's page.
• Lower the weight when faces or hands start to warp.
• Use the same seed to compare weights fairly.
• Match the LoRA to its base model; an SDXL LoRA will not load on MOSS-Image.

MOSS-Image is a new open-source model from the OpenMOSS team that follows long, detailed prompts and renders English and Chinese text. Its community is training LoRAs for it quickly, and Mage carries a growing set of them: painterly and anime styles, ink and charcoal looks, lighting and detail sliders, and character LoRAs. See the MOSS-Image model page for its strengths and prompt tips.

Because MOSS-Image reads long prompts closely, a style LoRA and a carefully described scene work well together: the LoRA supplies the look, and the prompt controls what is in the frame and any words that appear in it.

On the Import page, paste a CivitAI model link and Mage recognizes the base model, downloads the file, and checks that it fits. Imports finish in about a minute, then appear in the LoRA picker for that model. Standard LoRA and LoCon files work; MOSS-Image does not take LoKr, LoHa, or DoRA files yet.

Importing is part of the Pro Plus and Max plans. On any plan you can use the LoRAs already in the catalog, and the unlimited AI image generator page explains which models each plan runs without limits.