Skip to content

Learn

What Is a LoRA? How It Works

A LoRA (Low-Rank Adaptation) is a small add-on file that teaches a large AI model a new style, character, object or concept without retraining the whole model. It stores compact weight changes that are added on top of the base model at generation time, with a strength setting that controls how strongly the effect shows.

Start on the free Starter plan.

LoRA - FunAI
LoRA - FunAI

A Small File That Steers a Big Model

The base model stays exactly as it is. A LoRA adds a compact, adjustable correction on top, which is why one model can switch between many styles and characters just by loading different LoRAs.

Try FLUX 2 LoRA Gallery Realism

Frozen base model

The original weights never change, so the model keeps everything it already knows.

Low-rank update

Two thin matrices approximate the change you want, which keeps the file small and quick to train and share.

Adjustable strength

A scale value blends the LoRA in lightly or strongly, and several LoRAs can be mixed in one image.

LoRA-ready models on FunAI

Some models come with a LoRA already loaded; others accept LoRAs as an input. Each model page lists its inputs and its price per run.

  • •
    FLUX 2 LoRA Gallery Realism:FLUX.2 with a Realism LoRA for natural skin, light and camera detail, plus a Lora Scale input for strength.
  • •
    FLUX 2 LoRA Gallery HDR Style:FLUX.2 with an HDR Style LoRA for vivid, high-dynamic-range looks.
  • •
    Qwen Image:Accepts a LoRAs input and merges up to three LoRAs into one generation.
  • •
    HiDream I1 Full:Open-source text-to-image model with a LoRAs input, where each LoRA has its own path and scale.
  • •
    FLUX.1 [dev] with Controlnets and Loras:FLUX image-to-image with LoRA, ControlNet and IP-Adapter support for precise style and structure control.
  • •
    MiniMax H3 LoRA:A video model that loads a trained LoRA at an adjustable strength and generates video with synchronized audio.
Browse all image models
LoRA Models

LoRA explained

Base image models such as FLUX, Stable Diffusion or Qwen Image know a huge range of subjects, but not your studio's illustration style or a specific recurring character. Fine-tuning the whole model to teach it one thing is slow, needs serious hardware and produces another multi-gigabyte copy. LoRA, introduced by Microsoft researchers in 2021 for language models and quickly adopted for image models, is the lightweight alternative.

A neural network is mostly large matrices of weights. LoRA freezes those weights and, for selected layers (typically the attention layers), learns a pair of thin matrices whose product is a small, low-rank update to the original. Because the rank is small, often somewhere between 4 and 128, the file is usually megabytes rather than gigabytes. At generation time the update is multiplied by a scale and added to the base weights: a strength of 0 means no effect, and higher values push the output further toward what the LoRA learned. Several LoRAs can be applied together, each with its own scale.

Use a LoRA when a prompt alone cannot reliably produce something specific: a house illustration style, a recurring character, a product, a camera look such as HDR or film realism, or a narrow skill such as re-rendering a scene from new angles. A LoRA only works with the base model family it was trained for. A FLUX LoRA will not load into a Stable Diffusion model, and a LoRA trained on an older version of a model may behave differently on a newer one.

Tips for working with LoRAs

  • Use the trigger word if the LoRA has one; many are trained to activate on a specific token in the prompt.
  • Match the base model and version the LoRA was trained on, or expect weak or broken results.
  • Stack carefully: two style LoRAs at full strength tend to fight, so lower each one when you combine them.
  • Reach for a LoRA when you will reuse a look many times; for a one-off character, a reference-image Tool is quicker.
  • Check the licence of any LoRA you download, since it carries terms from its creator and from the base model.
  • Record the LoRA, scale, seed and prompt of every result you like, so you can reproduce it later.

How to Use a LoRA on FunAI

Feature visual preview 1
Feature visual preview 2
Feature visual preview 3
Feature visual preview 4
01

Start with a LoRA that is already loaded

The FLUX 2 LoRA Gallery models (Realism, HDR Style, Digital Comic Art) each come with a LoRA built in, so you only need to write a prompt.

02

Set the LoRA strength

These models have a Lora Scale input. Start at the default, lower it if the style overwhelms your subject, and raise it if the effect barely shows.

03

Bring your own LoRA

Qwen Image, HiDream I1 Full, Stable Diffusion 3.5 Large and FLUX.1 [dev] with Controlnets and Loras accept a LoRAs input, where each entry gives the LoRA file path and its scale.

04

Compare with and without

Run the same prompt and seed with the LoRA off and on. Any difference you see comes from the LoRA, which makes it easy to pick the right strength.

No public creations to show yet.

Everything you need to know