Disclaimers 免责声明

⚠ The model released is intended solely for academic exchange and personal learning purposes, and must not be used for any commercial activities. As the author of this model, I will not be held responsible for any copyright or other legal issues that may arise from the use or misuse of the model by any individual or entity. By downloading or using this model, you agree to the above terms. Any legal issues that result from the use of this model are the sole responsibility of the user.

⚠ 本模型仅供学术交流和个人学习使用，不得用于任何商业目的,我作为该模型的作者，不对任何个人或集体使用该模型可能引起的版权或其他法律问题负责,下载或使用该模型即表明您同意上述条款,若使用该模型时引发任何法律问题，一切后果由使用者自行承担,

Intro. 简介

⚠ This model prohibits any commercial use !!! This model is for learning and communication purposes only !!! ⚠

⚠ 该模型禁止任何商用 !!! 该模型仅供学习交流 !!! ⚠

⚠ If you believe that your rights have been infringed upon, please contact me directly to delete it !!! ⚠

⚠ 如果您认为权益受到了侵害请直接联系我删除 !!! ⚠

---

? 2024/1/19

? There is no good or bad distinction between V1.0 and V1.0-RLHF50, it all depends on your personal preferences !

? V1.0和V1.0-RLHF50之间没有好坏之分，这完全取决于您的个人喜好！

---

? It is recommended to use the A5 Stabilizer for v1.0 (not the RLHF50 ver.) to see slightly different styles under the same random seed, and in some cases, improve the image to better match your preferences.

? v1.0 建议搭配 A5稳定器 来使用 (RLHF50 ver. 不要使用)，这样可以在相同随机数种的情况下查看略微不同的风格，并且可以在某些情况下改善画面让其更符合您的偏好,

? Recommended Inference Settings 推荐的推理设置:

CFG: 5.5~7

Sampler: Euler a

Steps: 28~35

High-res.: RealESRGAN_x4plus_anime_6B, 15 steps, 0.55~0.6 denoising strength

? This model is merged from 2 self-trained A5 checkpoint (trained by 12 and 13 images, respectively). Yes, that's all of the training set, which shows the potential of full-parameter fine-tuning. The diversity of this model is trade-off for clear lines and bright colors.

? This model has serious overfitting issues in some cases. That's why you have a small chance of directly seeing the content of the training set,

? 该模型由2个自训练的A5检查点（分别由12个和13个图像训练）合并而成,是的，这就是训练集的全部内容，它显示了全参数微调的潜力,该模型的多样性被交换为更鲜明的线条与明亮的色彩,

? 该模型在部分情况下有严重的过拟合问题,正因如此，您会有小概率直接看到训练集的内容,

? The author strongly dislikes many homogenized styles nowadays, so the author trained and shared this model.

? 作者非常讨厌现在许多同质化的风格，因此作者训练并分享了这个模型,

描述:

Update Note

? 2024/1/19 v1.0-RLHF50 is released. This is a checkpoint which is trained by RLHF based on v1.0 model for 50 epochs (200 steps, 3200 comparisons). The training method is same to A5 Stabilizer.

? 2024/1/19 v1.0-RLHF50发布。这是一个基于v1.0版本训练的模型，使用RLHF训练，方法与A5 Stabilizer相同。

? Initially, the author hope that RLHF would reduce the volatile performance of v1.0. However according to observation, interestingly, this version produces much rougher lines.

? 最初，作者希望RLHF能提升v1.0版本的稳定性，使得两个训练的画风能够好的融合。然而，根据观察，有趣的是这个版本产生了更粗糙的线条。

训练词语: