IMPORTANT — PIXAI REDISTRIBUTION PROHIBITED
DO NOT upload, repost, copy, redistribute, merge, convert, or publish this model — or derivatives intended to reproduce it — on PixAI.art.
This model is already officially available on PixAI.art through my own creator profile.
Any unauthorized redistribution, re-upload, conversion, or publication of this model on PixAI.art is strictly prohibited.
Please respect the original release and use the official PixAI version when using this model on that platform.
About this model
This is an Anima-based DiT model focused on delivering a strong, consistent and recognizable anime aesthetic.
The main goal of this model is not simply to produce anime-looking images, but to maintain a solid visual identity across different characters, compositions and scenarios, with clean rendering, expressive characters and a stable illustration style.
Compared with more neutral Anima checkpoints, this model is intentionally biased toward a polished anime appearance, making it particularly suitable for character illustrations, key visuals, portraits, dynamic scenes and anime-style artwork.
Key Features
Strong and consistent anime style
Clean and well-defined linework
Rich colors and solid shading
Strong facial rendering
Strong response to composition, pose and scene descriptions
Suitable for both simple character prompts and more complex scenes
Supports Danbooru / TAG-style prompting
Supports natural-language prompts
TAGs and natural language can also be mixed in the same prompt
Designed to achieve good results with a relatively low number of inference steps
Being based on the Anima / DiT architecture, the model benefits from a modern image-generation architecture with strong prompt interpretation and good scene coherence.
The model works particularly well when the prompt clearly describes the subject, appearance, clothing, pose, composition, environment and lighting rather than relying only on large amounts of generic quality tags.
Character Tags — AnimaDex
Anima includes knowledge of a very large number of characters from anime, games and related media.
To find the correct character tags supported by Anima, you can use:
AnimaDex is especially useful for finding the exact tags required to generate characters already known by the model.
Using the correct character tag is strongly recommended when trying to reproduce a specific character, since it allows the model to make use of the character knowledge already present in Anima rather than describing the character entirely from scratch.
Recommended Settings
Option 1 — Recommended Default
Sampler: Euler
Scheduler: Simple
Steps: 12–15
CFG: 3–4
This is my recommended starting point and provides a good balance between image stability, detail and generation speed.
Option 2 — Alternative
Sampler: ER SDE
Scheduler: BETA
Steps: 12–15
CFG: 3–4
ER SDE + BETA can produce a slightly different interpretation of details, linework and composition and is worth trying depending on the type of image you are generating.
Prompting
The model can be prompted using traditional anime TAGs:
masterpiece, best quality, 1girl, long black hair, blue eyes, school uniform, standing, city street, sunset, cinematic lighting
or with natural language:
An anime girl with long black hair and blue eyes is standing on a quiet city street at sunset. She is wearing a Japanese school uniform, while warm sunlight creates long shadows and soft orange reflections around her.
You can also freely combine both approaches.
For example, you can use character and appearance TAGs while describing the scene, composition and lighting in natural language.
For complex scenes, descriptive prompts are generally recommended. Clearly specifying the character appearance, pose, camera angle, environment and lighting usually gives the model more control over the final composition.
For known anime or game characters, check https://animadex.net/ to find the appropriate character tag before building the rest of the prompt.
Notes
CFG values around 3–4 are recommended. Increasing CFG too much is not necessarily beneficial and may make the image look overly forced or reduce the natural balance of the rendering.
The model has been optimized to already produce strong results at 12–15 steps, so very high step counts are generally unnecessary.
As always, sampler behavior can vary depending on the prompt, resolution and LoRAs being used, so feel free to experiment around the suggested settings.
If you are looking for a model with a strong anime identity, clean rendering and modern DiT prompt understanding, this checkpoint is designed exactly for that.