DiT Models
PixAI's newest model generation — Tsubaki and Serin — built to understand natural language like a person.
DiT models (Diffusion Transformer models) are PixAI's newest generation of in-house models. Where older models want comma-separated tags, DiT models actually understand full sentences — describe a scene the way you'd tell a friend, and they get it: who's where, what they're doing, and how the scene should feel.
The DiT models on PixAI
Tsubaki — the flagship anime model. Its standout strengths: keeping the same character consistent across scenes and expressions (a comic creator's dream), handling multi-character scenes without blending people together, cinematic compositions, and real-world knowledge.
Serin — authentic Korean manhwa/webtoon style: refined faces, luminous expressive eyes, vibrant layered colors, and equally strong male and female characters.
How to prompt them
Skip the tag lists. Write a flowing description — subject, appearance, action, environment, and mood:
> "A blushing schoolgirl nervously adjusts her glasses while clutching a love letter, cherry blossom petals dancing around her as she stands outside the classroom door."
Aim for roughly 30–150 descriptive words, or just leave Prompt-Helper on: it converts anything — bare tags, your native language — into the descriptive style DiT models understand best.
DiT LoRAs
DiT models support their own LoRAs (regular and XL LoRAs won't work — see Base-Model). You can train DiT LoRAs on PixAI; training takes about 70 minutes, and the multi-version system lets one LoRA support several base models. See LoRA-Training.