MoeEye/z-image-illustration-enchance

这是一个基于Z-Image Base进行微调，以精心挑选的10万张多样风格的插图作为训练集，尝试改善原模型生成动漫/二次元图像效果中存在AI观感及光影的模型。

因为是第一次做全参数finetune，在数据集准备以及训练参数上还有很多可以优化的空间。

已知问题有如下几个：

1.偶尔出现的手部和肢体缺陷，这在原底模中也会少量存在；

2.对于背景中复杂细节会出现扭曲和色块情况，大概是由于学习率偏高导致；

3.偶尔发生生成的有效图像仅占画布的一部分，一般可用负向提示词进行规避。

推荐CFG为5，采样器使用res_multistep 或者 euler，调度器采用simple 或 normal。推荐提示词包括“二次元数字动漫插画风格”。

测试时使用的负向提示词：blurry, ugly, bad, text, 错误的高光和光影，错误的人体结构, 变形的手，错误的手部， chibi, Q版, missing finger, extra finger，多余的脚趾，变形的脚，破损的，模糊的，未填满画布，文字海报

This model was fine-tuned from Z-Image Base using a curated dataset of 100k illustrations with diverse styles to optimize ability to reduce ai-generated feel in anime illustration style images. As this is my first attempt at finetuning, there remains lots of point for optimization regarding both data preparation and training parameters.

Known Issues:

Occasionally, defects occur in hands or limbs (these also appear slightly within the original base model).

Distortions and color block artifacts may arise when rendering complex details in backgrounds; this is likely due to a high learning rate.

Sometimes generated elements focus on only part of the image. This can usually be avoided by using appropriate negative prompts.

Recommended Settings: CFG Scale: 5

Sampler: ResMultistep or Euler

Scheduler: Simple or Normal

prompt include "digital anime-style illustration"

negative prompts I use: blurry, ugly, bad, text, 错误的高光和光影，错误的人体结构, 变形的手，错误的手部， chibi, Q版, missing finger, extra finger，多余的脚趾，变形的脚，破损的，模糊的，未填满画布，文字海报

——————————————————————————————————————

少量测试效果如下：

上为Z-Image Base原模型，下为本模型，以较为复杂的提示词以及高对比度和多色彩进行测试。

The image at the top shows Z-Image Base (the original base model), while the bottom one is this finetuned version. They were tested using relatively complex prompts involving high contrast and multiple colors.

多风格测试：

上为Z-Image Base原模型，下为本模型，以同样元素但不同风格的提示词进行测试。

Top row displays Z-Image Base (Original Model), Bottom section features this finetuned version; both were tested with same elements but different style prompts.