For the complete documentation index, see llms.txt. This page is also available as Markdown.

LoRA Training

Learn how to train your own LoRA models on SeaArt AI. Covers image, editing, and video training workflows, dataset preparation, tagging, parameters, and testing.


🤔What Is LoRA

LoRA (Low-Rank Adaptation) is a lightweight fine-tuning method that teaches a base model a new character, style, or motion without changing the base model's weights. A LoRA file is small, loads on top of any compatible checkpoint, and can be dialed up or down with a weight slider - that is what makes your results controllable.

👇Which Training Type Do You Need?

SeaArt supports three LoRA training tracks. Pick the one that matches your goal:

  • Image Training: teach a character, object, or visual style from 20–40 still images. Best for portraits, art styles, and specific subjects.

  • Edit Model Training: teach an instruction-driven edit from 10–50 source/result pairs, so the model learns "do this to the image."

  • Video Training: teach a motion, effect, or video style from 4–10 short clips.

Training type
What you train
Data required
Guide

Image Training

Character, object, or visual style

20–40 images

Edit Model Training

Instruction-driven image editing

10–50 source/result pairs

Video Training

Motion, effect, or video style

4–10 clips (16fps, ~5s)


⚙️The General Workflow

Every track follows the same five steps. No setup or coding is needed — training runs in your browser, and most LoRAs finish within an hour.

1

Prepare your dataset

Collect 20–40 clean, watermark-free images that share the look you want to teach.

2

Preprocess and tag

SeaArt crops your images and writes captions for you; you just review and fix them.

3

Set parameters

Pick a base model, then keep the recommended defaults. Each child page lists the exact values.

4

Monitor training

Watch the preview images update, and stop early once they look right.

5

Save and test

Save the version you like best, then generate a few test images to confirm it works.


💡Before You Start

Before you start, make sure your material matches your base model's style, your dataset is clean and watermark-free, and you have a clear trigger word. Varied, high-quality data matters far more than volume. If this is your first LoRA, begin with Image Training.

Quality beats quantity. Adding weak or off-style images lowers the final model quality more than it helps.


📃Track Guides

Image TrainingEdit Model TrainingVideo Training

Last updated