Moss:RAEL - Moss: Reconstructing Anime's Expressive Linework

An anime-focused fine-tune of Moss.

heavily inspired by - https://huggingface.co/RicemanT/MossAD-Demo

Usage

learn to use this first: https://huggingface.co/Comfy-Org/MOSS-Image

Workflow

in this link right here:

https://huggingface.co/sselpah/MossRAEL/resolve/main/mossrael_workflow.json?download=true

it's very much like the default except ModelSamplingAuraFlow shift node set to 8.0, don't go lower than 6

Works nicely on 1mp-1.5mp

  • Full artist & anime in files under dataset_image_counts.json (obviously artists with 1-10 images were probably not trained that well)
  • To prompt artist use Art by @<artist name>
  • To prompt anime screencap use Art by @anime screencap
  • To prompt specific copyright use [Copyright: <copyright>];
  • No underscores needed
  • You are invited to try other prompting methods

Example:

A vertical three-quarter portrait of cinderella (nikke) standing beneath the awning of a closed flower shop at night. She has very long silver hair, blue eyes, and small blue star-shaped hair ornaments. She wears a glossy black motorcycle jacket zipped over a white high-collared shirt, tailored black trousers, and black boots. Her right hand holds the handle of a clear umbrella, and her left gently cups a small glass apple at waist level. She looks towards the viewer with a quiet, contemplative expression. Raindrops bead on the jacket and umbrella, while a soft blue shop sign and a warm streetlamp create contrasting reflections. Buckets of pale flowers are visible through the window behind her.

Art by @jindogod.

[Copyright: goddess of victory: nikke];

Examples

hoodjamesasuna 100kanojoscake yvonnedoingsomething whatsherfacesnow
tanigawarurirock seijururu ringekokafka ninomikubread
mumumuwkazusa massiveyognimi maanufurina jindogodcinderella
jindogodbehemoth gabiranoshinoko flippygalbrena andomarine

Training

  • Dataset: Custom closed dataset of 68,000~ danbooru illustrations/anime screenshots (85%/15%)
  • Preparation: All images captioned using Qwen 3.8 27B because I'm poor with no connections to NSFW Astra
  • Training configuration: AdamW8Bit, 80 batch x 2 grad acc, 1e-5 cosine, 20 epochs (1.337M~ images seen), 1024 varlen (less than 1% was under 1024), trained off Moss Image Base.
  • Compute: 8xH100 - total cost 1k$

Credits

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for sselpah/MossRAEL

Finetuned
(44)
this model