Anime Illust Diffusion XL

新增了上百种画风。我尝试了新的训练参数和训练策略，为此花费了大量的时间和金钱。这里是一些进展：(i) 我开发了一种简单而独特的的模型融合算法用于稳定模型构图，并将它应用在了 v0.52 和 v0.6 版本的制作上；(ii) 我开发了一套训练策略，用于解决如多类别数据训练中的数据不平衡等问题，您可以在模型介绍的英文部分中找到关于它的细节；(iii) 我用很大的训练集测试了一些前沿的训练参数，例如金字塔噪声等。可惜的是，我仍未得到良好可用的结果。New styles have been added.I spent a lot of time and money trying new training parameters and training strategies. Here is the progress:(i) I developed a simple but useful model merging algorithm to stabilize image composition. I applied it to make v0.52 and v0.6 versions;(ii) I developed a training strategy to solve problems such as data imbalance in multi-category data training, you can find details in the English part of the model introduction;(iii) I tested some advanced training parameters, such as pyramid noise, etc., with a relatively large training set. Unfortunately, I didn't get any good or usable result.

animeIllustDiffusion_v06.safetensors

v0.6

v0.8

v0.71

v0.7

v0.61

v0.6 Beta

v0.6 Alpha

v0.52

v0.51

v0.5

v0.5-alpha

v0.41

v0.4

v0.3

v0.2

v0.1

Files with SHA256: 3137f0f2c6804a0106a8bbe320c7b4d0fa41439c1adc91f1dc9eab8a394d9832