Train LoRA on PixAI: Complete Creator’s Guide (Character, Style & More)

▸ VIDEO_GUIDE

▶ Watch: PixAI LoRA Training Walkthrough on YouTube

Visual companion to this written guide → opens in new tab

▸ PRO_TIP

Can’t fully remove text? Add watermark, text to your negative prompts during generation. But cleaning beforehand is always better. You can also use Flow Edit to remove watermarks instantly — no Photoshop skills needed.

▸ EXAMPLE

Castorice | Honkai Star Rail (Haruka)

▸ THE_GOLDEN_RULE

Always train on the base model you regularly use for generation. If you typically generate with Haruka V2 → train your LoRA on Haruka V2.

▸ DEDICATED_GUIDE

Training DiT LoRAs requires different settings, dataset prep, and compatibility considerations.

For complete instructions on DiT-specific training — including supported PixAI DiT models, optimal parameters, and dataset recommendations — read our dedicated guide.

▸ READ_DIT_LORA_GUIDE

▸ COMPATIBILITY_CHECKLIST.md

[ ] Architecture: SDXL · SD 1.5 · DiT

[ ] Model family: Illustrious · Noob · etc.

[ ] Creator recommendations from docs

[ ] Successful community LoRAs on this base

▸ TEMPLATE

Character_name, permanent_features, optional_tweaks, [flexible_elements]

▸ EXAMPLE — CASTORICE

Castorice/hsr, long purple hair, low twintails, purple eyes, hair flower, pointy ears, crown of thorns, black tiara

⚠ WHY_AVOID_OUTFIT_DETAILS

Bake clothing into trigger words and you’ll need to delete and replace them every time you want a different outfit. This kills your LoRA’s reusability.

💰 50% OFF — DATASET_REUSE

Reuse the same dataset to train a new version and you’ll get a 50% discount on training fees. Iterate aggressively without breaking the bank.

▸ Q_01

What is LoRA training?

LoRA training teaches an AI model a new concept — a specific character, art style, pose, or outfit — so it can reproduce that concept consistently during generation. Once you train LoRA on PixAI, you can apply it on top of any compatible base model to recreate the concept on demand. LoRAs are lightweight add-on files (typically 10-200MB) that modify a base model’s behavior without retraining the entire model.

▸ Q_02

How many images do you need to train a LoRA on PixAI?

Most successful LoRA trainings on PixAI use 10 to 30 images. The exact number depends on your goal: character LoRAs typically need 15-30 images covering different angles, expressions, and lighting; style LoRAs work best with 20-30 images that consistently demonstrate the target aesthetic; pose LoRAs can succeed with as few as 10-15 images if the pose is well-isolated. Quality matters far more than quantity — 20 sharp, varied images outperform 100 blurry or repetitive ones. Minimum image resolution is 512×512 for SDXL bases and 768×768+ for DiT bases.

▸ Q_03

What is the best base model to train a LoRA on?

Always train a LoRA on the base model you regularly use for generation. A LoRA trained on Haruka V2 will work best when used with Haruka V2 — switching to a different base often degrades the result. The base model determines both the LoRA’s style understanding and which models it can be used with later. For modern DiT-quality workflows, train on Tsubaki (DiT.1) or Tsubaki.2 (DiT.2). For high-detail anime, the Illustrious family (Haruka, Hoshino, Otome) is the strongest SDXL choice. For vibrant, dynamic compositions, pick the Noob family (Hinata V2, Hikari).

▸ Q_04

How long does LoRA training take on PixAI?

LoRA training time on PixAI depends on the base model architecture. SDXL-based LoRAs (Haruka V2, Hoshino V2, Otome V2, Hinata V2) typically complete in 15-30 minutes. DiT LoRA training on Tsubaki or Tsubaki.2 takes approximately 70 minutes due to the larger model architecture. The initial time estimate shown during training may be less accurate at the start, so plan around the longer end of these ranges. You can check progress anytime in your Profile → Models/LoRAs tab.

▸ Q_05

Can I use the trained LoRA for commercial use or sharing on social media?

Yes. You can use your LoRA for commercial projects, commissions, and social media posts. However, if your LoRA is trained on existing characters or copyrighted materials, please follow the rights holder’s guidelines. Commercial usage may vary depending on the copyright of your training data. Original characters and styles you created yourself face no such restrictions.

▸ Q_06

Can I train a LoRA on Tsubaki.2?

Yes. PixAI supports LoRA training on Tsubaki.2 as a standalone DiT.2 base model. Tsubaki.2 LoRAs use longer trigger phrases (recommended 30+ characters describing core visual features) compared to the short tags used in SDXL workflows. Training time on Tsubaki.2 is approximately 70 minutes. Important: LoRAs trained on any other model (Tsubaki, Tsubaki v1.1, Tsubaki Flash, Serin) will not work on Tsubaki.2, so you must train against Tsubaki.2 specifically. The dataset reuse feature lets you re-apply an existing dataset to Tsubaki.2 with a 50% training discount.

▸ Q_07

Can I update or retrain my LoRA later?

Yes. You can create new versions of an existing LoRA with improved datasets or refined trigger words, without losing your existing followers or stats. If you reuse the same dataset to train a new version, you’ll receive a 50% discount on the training fee. All versions are managed under the same LoRA model card. See our Multi-Version LoRA Guide for the complete workflow.

▸ Q_08

What’s the difference between a LoRA and a base model?

A base model (also called a checkpoint) is a large, self-contained AI system trained on massive datasets — it can generate images on its own. A LoRA (Low-Rank Adaptation) is a small add-on file that modifies a base model’s behavior for specific characters, styles, or poses. Think of the base model as an engine and the LoRA as a custom tuning chip. LoRAs are typically 10-200MB versus multi-GB for base models, train in minutes-to-an-hour instead of days, and require only 15-30 images instead of thousands. On PixAI, you always load a base model first; LoRAs are optional add-ons stacked on top.

▸ Q_09

Is LoRA training free on PixAI?

LoRA training requires credits on PixAI, but Membership plans include free monthly LoRA training quotas. The Starter plan includes 3 free LoRA trainings per month, Hobbyist includes 5, and higher tiers include more. If you reuse a previously uploaded dataset without modifications, you get an automatic 50% discount on training fees. Free-tier users without membership can still train LoRAs by spending earned daily credits — there’s no hard paywall.