A new LORA has arrived ?!

OKAY GUYS FIRST OF ALL LETS ADDRESS THE ELEPHANT IN THE ROOM I KNOW ITS BEEN A VERY LONG WAIT AND YOU DID NOT EXPECT A LORA, SORRY FOR THAT !

More seriously tho, I add some big problems and cut off myself from the internet... Now, I pace everything without giving me any pression, so expect not scheduled release of some sort (could be next days, weeks, months, etc...).

I want to thank you all guys for the supports you gave for my models (I started this as a hobby of some sort), and it grew far beyond my expectations ! Love to it, and its all thanks to you guys <3.

AlcAI - Animehaven | MOSS-Image Anime LoRA

A LyCORIS LoKr for MOSS-Image-Base that brings high-quality anime art

(gacha card art, detailed high-res fanart) to the base model. It was trained on

anime images in many different styles, so it does not lock you into one look. Describe the style you want in the prompt and

the LoRA renders it as anime.

Recommended settings

Settings and Values

LoRA strength : 0.5 to 1.6 (start around 1.0)

Base model : MOSS-Image-Base (OpenMOSS-Team)

Sampling steps : 50

CFG : 3.0

Shift : 5.0

Resolution : about 1 MP (for example 832 x 1216 portrait)

Trigger word : none needed

How to prompt

Write natural-language prompts in two parts, the way the training captions were written:

1. Content: subject, pose, clothing, framing, background.

2. Style, in concrete words: line work (thin or bold outlines), shading (cel,

soft gradient, painterly), color palette, lighting, level of detail, mood.

"Anime style" alone gives a generic result. A described style, such as "bold black

outlines, flat cel shading, glowing yellow accents", gives you that style.

What to expect

• Clean anime rendering across a wide range of styles: illustration, flat graphic scenes, painterly fanart.

• It follows a style written in the prompt instead of forcing one look.

Training details

Type : LyCORIS LoKr (factor 8, full matrix)

Dataset : about 4,160 curated anime images, many style

Captions : per-image natural-language descriptions of content and style (Qwen3.5-35B-A3B)

Batch : 64 images per step, at about 1 MP each

Training steps : 2,000 |

Total training steps : (4,159 images / 64 batches) x 2,000 steps = about 130,000

Learning rate : 2e-3, cosine schedule, 100 warmup steps

Precision : bf16

Hardware : 1x RTX 5090 (32 GB)

Training time : about 31 hours

Dataset preparation :

• Images under 0.5 MP and extreme aspect ratios were removed.

• Each image got a caption describing its content and drawing style.