<h1 id="heading-6389">Please use <a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-webui-additional-networks">sd-webui-additional-networks</a> to load this model !</h1><h1 id="heading-6390">请使用 <a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-webui-additional-networks">sd-webui-additional-networks</a> 来加载本模型！</h1><hr /><h1 id="heading-6391">Introduction 简介</h1>Q0: What is this model? 这个模型是什么？A0: 🎈 This is a stabilizer trained for <a target="_blank" rel="ugc" href="https://civarchive.com/models/9409?modelVersionId=30163">Anything V5</a>. 这是一个为 <a target="_blank" rel="ugc" href="https://civarchive.com/models/9409?modelVersionId=30163">Anything V5</a> 训练的“稳定器”。Q1: What's the use of this model? 这个模型有什么用？A1.1: 🎉 This model can to some extent fix the collapsed structure in the generated illustration. 该模型能够一定程度上修复画面的崩坏部分。A1.2: 🎨 This model can fix the pupils of the generated characters. 该模型能够修复生成人物的瞳孔。A1.3: 😎 This model modifies the lightning effects of the generated illustrations, so that they will be less look like generated by AI (i.e., prefered by human). 该模型会调整生成画像的光影效果，使之看起来“没那么像是AI生成的” （换言之，被人类偏好）。A1.4: 🍡 This model seems to better align the generated image with the input prompts. 该模型似乎能够更好地对齐生成的图片与输入的提示词。Q2: How is this model Trained? 这个模型是如何训练的？A2.1: 🥰 This model is trained by Reinforcement Learning from Human Feedback (RLHF), which has been widely used in LLMs. 该模型由人类反馈强化学习技术训练，该技术已被广泛应用于微调LLMs。A2.2: 😏 To be specific, this model is trained by <a target="_blank" rel="ugc" href="https://github.com/yk7333/d3po">D3PO</a>, see <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2311.13231">Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model (arxiv 2311.13231)</a>. 具体而言，该模型的训练基于<a target="_blank" rel="ugc" href="https://github.com/yk7333/d3po">D3PO</a>方法，详见<a target="_blank" rel="ugc" href="https://arxiv.org/abs/2311.13231">Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model (arxiv 2311.13231)</a>。A2.3: 🙄 This model is instructed by a self-trained reward model in the training process, using 4 metrics (likes, collections, AI-generated probability, and views). The reward model is available at <a target="_blank" rel="ugc" href="https://huggingface.co/chikoto/ConvNeXtV2-IllustrationScorer">chikoto/ConvNeXtV2-IllustrationScorer</a>. 该模型在训练时由一个自己训练的奖励模型指导，使用 点赞数，收藏数，由AI生成的概率，以及浏览量 四个指标来评价。该奖励模型可以在 <a target="_blank" rel="ugc" href="https://huggingface.co/chikoto/ConvNeXtV2-IllustrationScorer">chikoto/ConvNeXtV2-IllustrationScorer</a> 找到。<hr /><h1 id="heading-147">Some Nonsense Crap 一些无关紧要的东西</h1>🧐 Any idea and suggestion is welcomed! 欢迎任提供何想法和建议！🧐🤗 Welcome to challenge NAI3 with a mortal body. 欢迎各位以凡人之躯挑战NAI3 :) 🤗

A5 Stabilizer - RLHF D3PO - A5稳定器

2024/1/6 First released the 210-epoch checkpoint.

Files with SHA256: 083a83adae7aab09bcda1855e45ce5efe4a15aca98cb7d0800b34eda1c426d96