La Luminette de la Nuit
Luminette, the smaller variant of the Lumina 2 checkpoint. I have torn off her tiny parts (diffusion layers) and glued together a new doll, carefully crafted her training regiment, and finally sent her to the wild.
For those who need help with metaphors:
Please, do not have high expectations. This is not the full model, but a smaller, unique one. The lite version was trained further in full precision. It was born as a playground for "knowledge transfer".
The training pipeline and the converter scripts were built from scratch.
Prior to training, data preprocessing and image alignment have been made to improve efficiency. Some of the complexity is being taken out, and the focus is on generating unique styles and features. This means that instead of trying to figure out how to draw the same thing from lots of different angles, the model can just render figures with their own distinct looks and personalities.
Please do not reupload to Huggingface, TensorArt, in the original, or in any smart compressed format. I can do it myself. Thank you.
The Lumina models keep their text encoder and vae unchanged, with only the diffusion model uploaded.
Description
Stabilized the training by removing the modulation layer from the loop. The model doesn't collapse, the images have less burn marks.
The abridged dataset was sampled from paintings, character illustrations. The smaller part has fashion items, hair and dress tags, body tags (eyes, feet), these images are with the correct body proportions.
Realigned, recaptioned the images, kept the most frequent tags. The automated preparation alone took days of compute time. Following the 80/20 rule, the result by effort and prompt understanding is similar as if the model were trained on all images. Meanwhile kept the dataset balanced.
Attempts were made to make the style closer to the most frequent tags; it cannot replicate characters.
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.