LoRA Training for Anime Characters: What to Prepare Before You Train

Published on October 23, 2026 by Steven Jones

When a character LoRA turns out badly, it’s tempting to blame the settings: too many steps, the wrong learning rate, or the wrong strength. More often, the problem is decided before training begins, by the images you use.

Many platforms now make LoRA training a few clicks. The real work happens beforehand: picking the right base model, collecting images that teach one clear concept, and choosing trigger words that leave room for variation. Here’s what to prepare for an anime character.

First, Do You Actually Need a LoRA?

A LoRA (Low-Rank Adaptation) is a small add-on file that teaches a base model a specific concept, such as a character, style, outfit or pose, so it can reproduce that concept on demand. It is the right tool when you will come back to the same character again and again: a webcomic cast, a VTuber, an OC you post every week.

A LoRA is often the wrong tool for a one-off. If you only need the same character in two or three images, use a reference image or detailed prompt instead. It’s faster. Modern reference tools can carry a design across a handful of images without any training at all.

A useful test: if you have, or can make, a dozen or more good character images and expect to need many more, training makes sense. Otherwise, start with references.

---

---

Choose the Base Model Before Anything Else

A LoRA is tied to its training base. Use the model you actually plan to generate with. A LoRA trained on one SDXL-based anime model usually works best on that same model, and LoRAs trained on one architecture will not work on another.

That compatibility gap matters. On PixAI, Tsubaki.2 is a standalone base model: LoRAs trained on earlier models won’t run on it, so you have to retrain a character specifically for it. Before you collect a single image, decide where this character will live.

How Many Images Do You Need?

Older architectures need fewer images than you might expect. Newer ones need more. PixAI’s published training guides recommend:

LoRA type Recommended dataset (SDXL bases) What to vary

• Character 15–30 images. Angles, expressions, lighting
• Style 20–40 images. Subject matter, while keeping one consistent technique
• Outfit 15–25 images. Angles, lighting and poses, with the outfit itself unchanged
• Pose 10–20 images. Viewing angle, ideally with different characters

For DiT-based models, PixAI recommends 30–100 images: below 30, the LoRA doesn’t generalise well; above 100, you mostly add training time.

In every case, 20 sharp, varied images beat 100 blurry or repetitive ones.

What Belongs in a Character Dataset

Keep one character in each image. This is the rule people break most often. In a group shot, the model cannot tell which traits belong to which person, so distinctive details, like unusual eye colour, get averaged out or copied onto everyone. Crop group images down to your character, or leave them out.

Use clear, high-resolution images. 512×512 pixels is the floor, while 768–1024 pixels is better for SDXL training. A model cannot learn detail that is not in the source.

Remove text, watermarks, and speech bubbles. Anything repeated in the dataset can reappear in generated images.

Crop small details up close. An iris pattern, earring, or hair clip may occupy too few pixels in a full-body image for the model to learn it. Crop those details out as extra images, and keep the crops at a consistent size so the model learns the content, not the size differences.

Vary what you want to change later. Show the same outfit in every image, and the LoRA may treat it as part of the character. If you want to dress the character differently later, include a few different outfits.

If Your Character Only Exists in Your Head

Many OC creators don’t have 20 drawings ready to use. A common approach is to generate the dataset: lock the design with a turnaround sheet, then generate many images of the character in different angles, expressions and settings, and keep only the ones that match the design exactly.

Be picky when curating. Every off-model image teaches the LoRA a slightly different version of the character. Also vary the backgrounds and lighting on purpose, because whatever repeats across the dataset, the LoRA will learn as part of the character.

If you use existing art, use your own or art you have the rights to. Training on another artist’s work or a copyrighted character without permission creates problems the moment you share the LoRA or use its output commercially.

Trigger Words: Permanent Features Only

The trigger word activates the LoRA. Describe permanent features only: hair colour and style, eye colour, markings, and signature accessories. Leave out clothing, poses and backgrounds. If an outfit is baked into the trigger, you’ll fight it every time you want a different look.

The right length depends on the model. On PixAI, LoRAs for Tsubaki (DiT.1) work best with short, focused triggers, while Tsubaki.2 (DiT.2) recommends trigger phrases of at least 30 characters that describe the character’s core features in detail. Check the guidance for your base model before writing yours; PixAI has a separate guide to trigger words.

What Training Costs in Time and Credits

On PixAI, SDXL-based LoRAs typically take 15–30 minutes. DiT LoRAs take an hour or more because the architecture is larger. The initial time estimate is the least reliable, so leave extra time.

Training costs credits, although membership plans include a monthly allowance of free training runs. Retraining on the same dataset gets a 50% discount, making iteration cheaper: train a first version, test it, and retrain with a better trigger or a newer base.

Your search, your sources

Make DIY AI a preferred source

See more of our reporting in Google Top Stories, AI Overviews and AI Mode.

Add DIY AI on Google

Test It Before You Rely on It

A LoRA that reproduces your training images is not finished yet. Test what you will actually ask of it:

• Check strength. Try the same prompt at several LoRA strengths. If the character appears only at the highest setting, the LoRA is undertrained, or the trigger is weak; if it looks right only at low strength, it may be overfitted to the training images.
• Try a different outfit. If the training outfit keeps appearing, the dataset repeated it too much, or the trigger slipped in.
• Change the pose and setting. Repeated poses or backgrounds in the output usually point to too much repetition in the dataset.
• Check close-ups and full-body shots. Small features need to hold up in both.

Each failure points back to a specific fix in the dataset, which is why the dataset matters more than the settings.

Where PixAI Fits

PixAI lets you train LoRAs in the browser without a local GPU, on its anime-focused SDXL models as well as its DiT models, Tsubaki and Tsubaki.2. Finished LoRAs load straight into the generator, can be kept private or shared on the model market, and can be updated as new versions without losing the original listing.

Its LoRA training guide covers dataset preparation, naming, and trigger words step by step, and the DiT LoRA training guide explains what changes the newer architecture introduces.

Prepare First, Then Train

Most LoRA work happens before training starts. Pick the base you will generate with. Collect 15–30 clean images of one character (30 or more for DiT), with variety where you want flexibility. Write a trigger made of permanent features. Then train, test against new outfits and poses, and retrain with what you learn.

For a character you return to regularly, build the dataset and try a first version on PixAI. The first LoRA is rarely the last one, but it is the one that shows you what your dataset was missing.