The art style of Astrae Oratio's official illustrations: a cool navy and teal palette, clean cel shading, soft light and four-pointed star sparkles.
Two versions, trained on the same 78 images: Anima (prompt with tags) and Krea 2 (prompt with plain sentences). Pick the one for your base model; how to prompt it and the settings are under About this version.
Trigger: astrae_oratio. It is a style, not a character: describe your own character after it. Weight 1.0 in every gallery image.
Each gallery is original characters in places the game art never shows, from a concert stage and a space station to a set of little witches at the beach, on a train and at a night festival, so it shows the style holding up away from what it was trained on. Each version's gallery also has a sheet of every scene without and with the LoRA (same prompt and seed), and a sheet of every render its gallery was picked from, so you can see the hit rate.
Training images
78 pieces of official Astrae Oratio art: key visuals cropped clear of the logo and text, cut-in illustrations, and character sprites placed on white. Captions describe everything in the picture with no words about the drawing style, so the trigger only has the style left to learn.
Limitations (both versions)
It leans toward the game's own look: an outfit you do not name can turn into a dark school blazer, and palettes drift navy. Name the outfit and the background.
Four-point sparkles show up even when you don't ask for them.
Credits
Astrae Oratio art belongs to its rights holders. Anima by CircleStone Labs, Krea 2 by Krea. Gallery checkpoints: JANIMA v10 (an Anima merge) and RedCraft Krea 2 v2.2. Anima samplers by KeithZ117 (AnimaFlow) and BlackSnowSkill (FLS).
Description
Anima LoRA, step 1350. Trained on Anima base v1.0 with WD14 tag captions (astrae_oratio first), rank 32 / alpha 32, lr 1e-4, 1024, 4 repeats, gradient accumulation 4, horizontal flip, 1350 steps (about 17 epochs), AnimaLoraToolkit.
Prompt with tags, trigger first:
astrae_oratio,
masterpiece, best quality, score_7, newest, safe,
1girl, solo,
miko, japanese clothes, white kimono, red hakama, long black hair, hime cut, hair ribbon, holding broom,
full body, standing, looking at viewer, smile, shrine, torii, stone stairs, autumn leaves, falling leaves,
maple tree, afternoon
Settings: the gallery uses JANIMA v10 with AnimaFlow (25 steps, CFG 5, shift 3) and an FLS refine pass (30 steps, denoise 0.7), CFGZeroStar, about 2.2 MP. The workflow is in the Archive file and in every gallery PNG. Custom nodes: KeithZ117/Comfyui-anima-sampler and BlackSnowSkill/ComfyUI-BSS_FLSampler. It also works on the official Anima base v1.0 (gallery image 21: same seeds and prompts).
Quirks: a plain scene can come out framed like a game sprite on white, so name a background. In scenes with two people, one is sometimes dropped. Only tested with the workflow above; a plain KSampler should work but is untested.














