⭐If you have the means, please do support the original artist's creative work!

A LoRA that reproduces the drawing style of the artist 鬼针草 (chen bin / szcb911).

Artist page: https://www.pixiv.net/users/6049901

Trigger word: @szcb911style — put it at the very beginning of your prompt

V2 CHANGELOG: Using KohakuMoss as the training tool, we readjusted the training parameters and training dimensions. v1 employed a training resolution of 1m (with the longest side being 1024), while v2 directly used an almost full-image resolution of 2m (with the longest side being 2048) for training. The RANK32 of v1 was upgraded to RANK64 in v2, and the DoRa format was adopted without adding new training layers. This approach enhanced the accuracy of composition while ensuring compatibility.

V1 TIPS: the stability of the foot in the generated image from the model is slightly unsatisfactory. I need to try several more seeds, The color is slightly different from v2

Base model: MOSS image base

Training data

200 images from the artist's posts with perform manual full-scale cleaning on the tags

中文

⭐如果您有条件，请支持原创艺术家的创作！

一个能够再现艺术家鬼针草（chen bin / szcb911）绘画风格的LoRA。

艺术家页面：https://www.pixiv.net/users/6049901

触发词：@szcb911style — 请将其置于你的提示词开头

换用KohakuMoss作为训练工具，重新调整了训练参数与训练尺寸
v1是1m（1024最长边）训练分辨率桶，v2则直接是2m（2048最长边）的几乎原图分辨率进行训练，v1的RANK32→v2的RANK64，并使用了DoRa格式，但没有增加新的训练层，在提升构图准确性的同时确保了兼容性。

v1提示：模型的肢体稳定性稍微有些问题，需要多抽几次seed，色彩与v2略微有些差别。

基础模型：MOSS image

训练数据

200张图片进行手动全面清理标签