CivArchive
    Preview 148715305
    Preview 148714908
    Preview 148715001
    Preview 148715017
    Preview 148715049
    Preview 148714764
    Preview 148715083
    Preview 148715122
    Preview 148715218

    This is my first LoRA ever! It's based on the art style of Astoro ( https://x.com/fider83304595 ), go check out their work. I couldn't find any LoRA for this style anywhere, so I figured I'd just make it myself. The dataset is mostly dragons and wolves, but it handles other animals and even humans pretty well.

    There are a few versions now, so please pick the one that matches your base model and check its own notes (the version tabs at the top). Every version uses @astoro at the start of the prompt.

    Krea 2 (v2, the new one)

    Retrained on my own PC on the Krea 2 Raw base. Epoch 14, strength 1.0, and write your prompts in full sentences like MOSS with @astoro at the start. The big fix over v1: it actually listens to clothes now. In my tests 101 of 104 images came out dressed as asked (v1 was about 1 in 10), and with just @astoro it gives the full style even when the base model wants to go realistic. Works great on Krea 2 Turbo and Neutrino. The full epoch comparison is in the version notes.

    MOSS (v2, the new one)

    Third try, now trained on my own PC with the same improved dataset as Krea 2 v2. Epoch 15, strength 1.0, and write your prompts in full sentences with @astoro at the start. Main fix over v1 the style is much closer to Anima now.

    Works with the MOSS Image Turbo LoRA ( https://civarchive.com/models/3006599/moss-image-turbo-loras?modelVersionId=3420002 ) [and pairs nicely with RDBT | MOSS for extra detail ( https://civarchive.com/models/3007506/rdbt-or-moss?modelVersionId=3457029 )]. The full epoch comparison is in the version notes.

    Anima

    Trained on my own PC with kohya on Anima Base v1.0. Works great at strength 1.0, you can go down to 0.7 if you want it softer. Works on Anima finetunes too and plays really nice with the Expressive & Accurate Turbo LoRA ( https://civarchive.com/models/2709753/expressive-accurate-turbo-for-anima?modelVersionId=3043931 ).

    Anima 2.9B

    The same LoRA remapped for the 2.9B models (they have 40 blocks instead of 28), no retraining needed. Only use it with 2.9B models. Works great at strength 1.0 too.

    NoobAI / Illustrious

    Trained on NoobAI-XL eps 1.1, so it works on most NoobAI merges and on Illustrious models too. Epoch 9, strength 1.0, regular tags with @astoro first. It looks great with a hires fix. It likes flat feet and can sneak purple into the background, check the version notes for tips.

    Krea 2 v1 (old)

    The first Krea version, trained on Civitai on the Krea 2 base. The style is nice, but it tends to go NSFW even when you don't ask for it. Please use the v2 (Krea 2) instead, this one is kept for the record.

    MOSS

    Round two after a horrible first try. Epoch 9, strength 0.8, and write your prompts in full sentences. Works with the MOSS Image Turbo LoRA ( https://civarchive.com/models/3006599/moss-image-turbo-loras?modelVersionId=3420002 ). It likes turning the character around, so say how he's facing in the prompt. It's decent, but honestly Anima nails this style better.

    v0.1 (old MOSS)

    My very first attempt, kept here for the record. You probably don't want this one.

    Description

    MOSS, round two: same improved dataset as Krea v2, trained locally

    v1 (MOSS) already behaved well, it followed clothing prompts fine, unlike Krea v1. But when I fixed the dataset for Krea v2, I wanted MOSS to learn from the same improved data so all versions stay consistent.

    What changed

    • Same 61-image dataset as Krea v2. That includes about 20 clothed (and shirtless-with-pants) images of the style, made with my own Anima version of this LoRA, captioned plainly as “fully clothed”.

    • Fixed captions on the original images, so the nude ones actually say so. The trigger learns the style, not the nudity.

    • Trained locally with sd-scripts on MOSS SFT, rank 32, alpha 32, lr 1e-4, 16 epochs (61 × 3 repeats × 15 = 2745 steps). About 4 hours on an RTX 5070 Ti. I went with epoch [15/16].

    The comparison
    I tested epochs 8 to 18 against a no-LoRA control, same seeds for all of them: hoodie and jeans at home, leather jacket in the city, hawaiian shirt at the beach, full plate armor in a castle, shirtless at a campfire, and a skinny rabbit accountant in a suit.

    • Clothes: [X of Y] came out dressed as asked.

    • Style: shows up around epoch [X] and is solid from [X].

    • Why epoch 15: looks the same as later ones without overcookin.

    • Respects the prompt

    How to use it

    • Natural sentences, @astoro at the start.

    • Strength [1.0].

    • Normal: [30 steps, CFG 4, eualer].

    • Turbo (8-step LoRA): works with [turbo LoRA name], 8 steps, CFG 1. At CFG 1 the negative prompt does nothing, so describe the clothes if you want SFW.

    • With RDBT | MOSS (v0.2): [tested / works well] together. RDBT is a detailer for lighting and textures with no style of its own, so it pairs naturally with a style LoRA like this one. Its author recommends strength 1 on MOSS base and lower on SFT. [Your result: e.g. “adds detail without washing out the style” + strength used].

    • For armor, describe the lower part too (chainmail skirt, tabard, plates covering the hips). [Remove if MOSS doesn’t have this problem.]

    FAQ

    LORA
    MOSS Image

    Details

    Downloads
    41
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/12/2026
    Updated
    11/22/2026
    Deleted
    -
    Trigger Words:
    @astoro

    Files

    Astoro_MOSS_v2.safetensors

    Mirrors

    CivitAI (1 mirrors)