LoRA Training
Train your own LoRA on PixAI in a few clicks — your character, your style, your rules.
Can't find the LoRA you need in the Model-Market? Train your own — no coding, no GPU, just your images and a few decisions. And once published, you earn Credits whenever someone else uses it.
Step 1: Prepare your dataset (this decides everything)
Your images are the ingredients — quality matters more than anything else:
How many: 15–30 for a character, 20–40 for a style, 10–20 for a pose (up to 100).
Resolution: at least 768px; blurry or tiny images produce blurry LoRAs.
One clear subject. For character LoRAs, avoid images with multiple characters — the AI won't know whose features to learn.
Clean images. Remove text, watermarks, and speech bubbles first (PixAI's Inpainting works great for this) — or your LoRA may learn to generate random text.
Step 2: Set up the training
Name it something searchable, e.g.
Castorice | Honkai Star Rail (Haruka).Pick the base model — the most important choice. Use the model you actually generate with; the LoRA will only work with models sharing that base. See Base-Model.
Set the trigger words — format:
character_name, source_work, attributes. Choose carefully: they can't be changed later. See Trigger-Words.
Training takes minutes for SDXL bases and about 70 minutes for DiT bases; you'll get a notification when it's done.
Nice extras
Multi-version LoRAs: add new versions of the same LoRA for different base models, all under one model card.
Dataset reuse: retrain with a previous version's dataset unchanged and get 50% off the training cost.