v0.96 日本文解説は英文以下に記載。モデレーターの審査通った・・💦

Since I wasn’t satisfied with the results generated by the NSFW prompts, I made some minor adjustments and additions to the tags and prepared three prompts with high generation rates, using as consistent a description as possible. When generating characters without LoRa, the prompt should end with a complete character description. When using LoRa characters, I recommend including the character trigger and a description of their appearance immediately after “1girl.” Please refer to the trigger word field and sample images for guidance.

v0.95　※The “branch lingerie” trigger tag has been discontinued in this version.

I reviewed the trigger tags, defined panties and bras separately, and adjusted the training parameters.
“small bare tree,” “〇〇 draped over a branch,” “〇〇 tangled in branches”
As a result, the final product offers both greater versatility and practicality.
As a specific example of its versatility, it can be used with bikini and school swimsuit outfits.
I don’t know who this benefits, but in the sample images, I’m trying out a challenge where I hang the Lora school swimsuit upside down.

When generating explicit NSFW content, perhaps because of compositional challenges, you need to run the batch to show both panties and a bra.

v0.93

日本文解説は英文以下に記載。

Now, I've been working hard to create a new concept piece to upload to Anima, but I’ve hit a wall once again. This reminded me of the early version of “bound to desk.” In short, since this Lora is still under development, please don’t expect too much when using this version.

Existing Issues

1.When generating images with people in them, the model tends to have the figures lean against branches, likely influenced by the word “draped.” Starting with an unreleased version, we avoided using the “hang” tag and selected alternative words in consultation with Grok, but this behavior remains unavoidable.

2.I want to add natural language captions using tags, but—likely due to an issue with the tool I’m using (AnimaLoRaFactory)—the tags in the generated LoRa training dataset were scattered across individual words. It appears that the tag dataset is not being generated using only tags consisting of up to four words, as required by the Booru language specification.

I did create an image dataset that includes people, but when I asked Grok, I was told that to ensure the model learns the concept itself accurately, it’s better to either exclude people entirely or limit their number to a few. Therefore, for this version, I trained the model using only the “no humans” dataset.

I’d also like to create a version trained with images that include people and compare the results.

さて、そろそろAnimaでのコンセプト新作をアップロードしようとしたところ、また壁にぶち当たってしまった。「bound to desk」の初期バージョンを彷彿とさせた。結論として、このLoraはまだ発展途上なので、このバージョンの使用に際して、あまりその生成結果を期待していないで貰いたい。
抱えている問題点
1. 人間を入れて生成すると「daraped」の単語につられて人間が枝によりかかる傾向が出る。未公開のバージョンから「hang」というタグの利用は避け、Grokに相談しながら別の単語を選定したが、それでも避けられない。
2. タグ付けで自然言語キャプションを入れたいが、使用しているツール(AnimaLoRaFactory)の問題なのか、できあがったLoRaのTraning dataset tagが単語単位でバラけていた。おそらくbooru語準拠の4単語までをタグとしたデータセットしかできていないみたいだ。

一応、人物を入れた画像データセットは作って有るが、Grokに聞いてみると、コンセプト自体の学習精度の為、全く入れないか、少数にとどめたほうがいいと言われていたので、本バージョンは人物が入っていない「no humans」のデータだけで学習させてある。
入れて学習させたものも作り比較したいと思う。