DiT LoRA Training on PixAI: Tsubaki, Tsubaki.2 & Serin Guide (2026)

DiT LoRA training has quickly become the most powerful way to create custom characters and styles on PixAI. With full support across both DiT.1 and the newer DiT.2 architectures, you can now train custom LoRAs against PixAI’s flagship DiT base models. This guide covers everything you need before your first DiT LoRA training run: which DiT models are available, which ones you can actually train against, expected training time, trigger word formatting (which differs between DiT.1 and DiT.2), recommended dataset specs, and the common pitfalls that trip up new trainers.

---

Why DiT LoRA Training Beats SDXL for Modern AI Art

When you generate scenes with two or more characters on SDXL-based models, you often run into “feature blending” — visual traits like faces, hair, and clothing start mixing between characters. The problem gets dramatically worse with three or more people in a frame.

DiT-based architectures handle multi-character composition far more reliably. They preserve each character’s distinct features, manage spacing and interaction more naturally, and produce coherent eye-lines, gestures, and overlaps. If you want a LoRA that holds up in busy multi-character scenes — or just want to take advantage of the newest base-model quality on PixAI — DiT LoRA training is now the better path.

---

All DiT Models on PixAI: An Overview

Here’s the full lineup of DiT base models available right now.

DiT.1 Models

👉 DiT-Based Models:

• Tsubaki

• Tsubaki v1.1

• Tsubaki Flash

• Serin

DiT.2 Models

Tsubaki.2 — PixAI’s flagship DiT.2 model. Substantially improved generation quality and prompt understanding over the DiT.1 line. 👉 Learn more here

---

Which DiT Models Can You Train a LoRA Against?

Not every DiT base model is exposed as a LoRA training target. As of 2026, PixAI supports DiT LoRA training against just two base models:

• Tsubaki (DiT.1)
• Tsubaki.2 (DiT.2)

The other DiT models — Tsubaki v1.1, Tsubaki Flash, and Serin — are available for generation but cannot be selected as a training base in the LoRA training panel.

That said, “trainable base” and “where your LoRA can run” are two different things. A LoRA trained on Tsubaki carries strong inference compatibility with Tsubaki Flash, which is exactly why Tsubaki remains the canonical DiT.1 training base on PixAI — train once, generate across compatible bases.

⚠️ Important: Tsubaki.2 is not backward-compatible Tsubaki.2 is a standalone base model. LoRAs trained on Tsubaki, Tsubaki v1.1, Tsubaki Flash, Serin, or any other model will not work on Tsubaki.2. To use a LoRA on Tsubaki.2, you must train it specifically against Tsubaki.2.

If you already have a successful Tsubaki LoRA and want to bring the same concept to Tsubaki.2, PixAI’s dataset reuse feature lets you re-apply the existing dataset in a couple of clicks — and applies a 50% training discount when the dataset is reused unchanged. Full workflow in our Multi-Version LoRA & Dataset Reuse guide.

---

DiT LoRA Training Time: What to Expect

DiT LoRA training takes longer than legacy SDXL training because of the heavier underlying architecture. Plan for:

• Tsubaki LoRA training (DiT.1): ~70 minutes per run
• Tsubaki.2 LoRA training (DiT.2): ~2 hours per run

The estimate shown in your training queue is a guideline, not a guarantee — actual duration varies with queue load and dataset size, and the early estimate is the least accurate. Queue the job and walk away rather than refreshing the page.

---

Trigger Word Best Practices for DiT LoRA Training

Trigger words on PixAI’s DiT LoRA system work differently from the comma-soup tag style many trainers remember from SDXL workflows. The rules also differ between DiT.1 and DiT.2 — this is the single most common place new DiT LoRA trainers slip up.

Trigger Words for Tsubaki LoRAs (DiT.1)

For Tsubaki LoRA training, keep trigger words concise and focused.

Character LoRAs — follow this format:

character_name, source_work, other_attributes

Example: hatsune_miku, vocaloid, twin_tails

Style or concept LoRAs — focus on the core shared characteristic of the dataset:

Example: ink_wash_style

General DiT.1 guidance:

• Keep trigger words short and specific
• Avoid long, multi-clause trigger phrases
• Identify the most essential common feature shared by every image in the dataset

Trigger Words for Tsubaki.2 LoRAs (DiT.2)

DiT.2 reverses the brevity rule. For Tsubaki.2 LoRA training, we recommend trigger words at least 30 characters long that describe the subject’s core features in detail. The richer the trigger phrase, the better Tsubaki.2 anchors your concept during training.

A strong Tsubaki.2 trigger phrase reads more like a compact feature inventory than a single token — distinguishing visual attributes, not just a name. If your trigger word is shorter than 30 characters, you’re leaving training quality on the table.

---

Recommended Dataset Specifications for DiT LoRA Training

Whichever DiT base you train on, the dataset rules are the same.

Image Count

30–100 images is the sweet spot for DiT LoRA training. Below 30, your LoRA won’t generalize; above 100, you mostly add training time without proportional quality gains.

Image Quality and Consistency

• Every image should share the common features you’re trying to teach (the same character, the same style, the same concept)
• Keep image dimensions consistent across the dataset where possible
• Aim for clear thematic or stylistic unity — a confused dataset produces a confused LoRA

Reusing an Existing Dataset

If you’ve already trained a LoRA on PixAI and want to retrain — for example, to bring an old Tsubaki concept onto Tsubaki.2, or to test a new training configuration — you can reuse your existing dataset directly. PixAI applies an automatic 50% discount on training cost when the original dataset is reused unchanged. Workflow details in the Multi-Version LoRA & Dataset Reuse guide.

---

How to Start Your First DiT LoRA Training Run

1. Go to PixAI’s LoRA training page.
2. Under Model Type, pick DiT.2 (to train against Tsubaki.2) or DiT.1 (to train against Tsubaki).
3. Confirm the Model Theme — Tsubaki for DiT.1, Tsubaki.2 for DiT.2.
4. Upload your 30–100 image dataset.
5. Set your trigger word following the rules above (concise for Tsubaki, ≥30 characters for Tsubaki.2).
6. Submit and wait. Tsubaki jobs take roughly 70 minutes; Tsubaki.2 jobs take roughly 2 hours.

---

DiT LoRA Training FAQ

Which DiT models can I train a LoRA on?

PixAI currently supports DiT LoRA training against two base models: Tsubaki (DiT.1) and Tsubaki.2 (DiT.2). Other DiT models — Tsubaki v1.1, Tsubaki Flash, and Serin — are available for generation but are not exposed as training bases in the LoRA training panel.

What’s the difference between DiT.1 and DiT.2 on PixAI?

DiT.1 is the original DiT generation on PixAI; the trainable DiT.1 base is Tsubaki, with Tsubaki v1.1, Tsubaki Flash, and Serin available for generation only. DiT.2 is the next-generation architecture and currently powers Tsubaki.2. DiT.2 produces higher-quality output, trains more slowly, and uses a different trigger-word style than DiT.1.

How long does DiT LoRA training take?

Tsubaki (DiT.1) LoRA training takes around 70 minutes per run. Tsubaki.2 (DiT.2) LoRA training takes about 2 hours. Times shown in the queue are estimates and can shift with system load.

Will a Tsubaki-trained LoRA work on Tsubaki Flash?

Yes. Tsubaki Flash is designed for high compatibility with LoRAs trained on the original Tsubaki, so a single Tsubaki LoRA generally works well across both bases at inference time. This is different from training compatibility — Tsubaki.2 remains incompatible regardless.

Can I use a Tsubaki LoRA on Tsubaki.2?

No. Tsubaki.2 is a standalone base model and is not compatible with LoRAs trained on any other model — including Tsubaki and Tsubaki v1.1. You’ll need to retrain the LoRA against Tsubaki.2. PixAI’s dataset reuse feature makes this fast and gives you a 50% discount on the retrain.

How long should my Tsubaki.2 trigger word be?

For Tsubaki.2 LoRA training, aim for at least 30 characters describing the subject’s core features in detail. This is a meaningful change from the short, comma-separated trigger words used for Tsubaki (DiT.1) LoRAs.

How many images do I need to train a DiT LoRA?

30 to 100 images is the recommended range for both Tsubaki and Tsubaki.2 LoRA training, with consistent quality, sizing, and clear thematic unity across the dataset.

Start Training Your DiT LoRA Today

Both Tsubaki and Tsubaki.2 LoRA training are live on PixAI right now. Pick the base that matches the look you want, follow the trigger-word rules for the generation you’ve chosen, and you’re ready to train.

Start Your DiT LoRA Training →