VRChat Binaco

Trained on animagineXLV3_v30, generate 3D model images of Binaco for sale on the booth, a new artist's attempt.

v0.1:

Recommended weight: 1.

Recommended ratio 1:1.

When using, it is recommended to include Animation style or use the Anime (动漫) option in the (recommended) SDXL Style selector.

The dataset consists of all 1024*1024 images, totaling 40.

Capable of generating outerwear and inner lining, but currently only one image is available for underwear.

During training, the tagger's tags were not altered, providing some degree of generalization. Current issues discovered include: changing only the hair color may affect the hairstyle, and there is a high probability of abnormalities in the generated jackets.

v0.2:

Compared to v0.1, the generated characters resemble more closely.

Recommended weight: 1.

v0.1 did not use adb buckets during training, while v0.2 used data in different proportions.

The dataset is trained at 1152*1152 resolution, comprising 72 images.

Utilizing the tagging method of trigger words + removing character's distinctive features tag. Trigger words for classic clothing: binaco coat. If the guiding effect is deemed insufficient, adding binaco on top is effective without trigger words.

==========================================================

基于animagineXLV3_v30训练，生成Booth上售卖的3D模型Binaco图像，新人尝试作品。

v0.1:

推荐权重：推荐1。

推荐比例1：1。

使用时推荐带上Animation style或者用 (推荐)SDXL Style样式选择器的Anime (动漫)

数据集全部都是1024*1024，有40张。

能出外衣和内衬，内衣目前数据集只有一张出不了。

训练时没有改变打标器的tag，有稍微的泛化能力，目前发现的问题有：只改变头发颜色可能会影响发型，出的外套很大概率异常。

v0.2:

较于v0.1，生成的人物更像了。

推荐权重：推荐1。

v0.1训练时没使用adb桶，v0.2使用了不同比例的数据。

数据集以1152*1152训练，有72张。

使用触发词+删除人物形象特征tag的标注方法。经典衣服触发词：binaco coat，不写触发词已经有效果了如果觉得引导力度不够再加上 binaco。