<h2>使用说明</h2><ul><li>禁止商用，如果在其他插画网站发布该模型生成的图片，请标记AI，我非常反感不打标记的行为。</li><li>推荐 DPM++ SDE Karras 20steps；Scale: 10 - 12；使用 ESRGAN_4x 放大，并且使用VAE得到更好的色彩。</li><li>这是一个绳艺特化模型，可能很多方面不如其他的模型。 最新的第三代版本支持口球(ball_gag)、眼罩(blindfold)，但是Sex Toy仍然画不好。</li><li>因为捆绑play插图中有非常多的细节需要学习，所以我进行了大量的训练，这可能会导致过拟合，使得泛化能力下降，不过可以通过模型融合来调整权重。</li><li>SFW和NSFW可控：你可以在Negative Prompt中加入"bdsm,nsfw,bound,bondage,blindfold,ball_gag,gagged,nipples,rope"等关键词，而且不能在Prompt中输入nsfw关键词来生成sfw图。欢迎各位大佬拿此模型融合！</li></ul><s>给个五星吧，爷真的求求你了</s><h2>Istructions</h2><ul><li>Commercial prohibited, If you post images generated by this model on other illustration sites, please mark AI, I am very much against unmarked behavior.</li><li>Recommended DPM++ SDE Karras 20steps; Scale: 10 - 12; use ESRGAN_4x to zoom in and use VAE to get better colours.</li><li>This is a rope art specialization model, which may not be as good as other models in many aspects.</li><li>The latest third-generation version supports ball_gag and blindfold. But Sex Toy still can't draw well.</li><li>Because there is a very large amount of detail to learn in the bundle play illustration, I did a lot of training, which may lead to overfitting and make the generalization less capable, but the weights can be adjusted by model fusion.</li><li>SFW and NSFW controllable: you can add to the Negative Prompt "bdsm,nsfw,bound,bondage,blindfold,ball_gag,gagged,nipples,rope", and you cannot enter the nsfw keyword in the Prompt to generate an sfw graph.</li></ul><s>If you like this model, please don't skimp on your 5 stars ;)</s>

Himitsu - HaneRu V3

V2.0：早上好，请慢慢赏玩，由啊雷洛炼制的模型。Info<ul><li>Base: AnythingV3 better VAE</li><li>Embedding is no longer needed</li><li>Better generalization ability</li><li>Enhanced the logic of the generating graph</li></ul>About ModelV2版本是一个融合模型，由于训练完的模型有些过拟合，故使用模型融合提高泛化。训练得到一个模型X，然后复制两份（记作X1，X2）分别与AnythingV3和OrangeMix模型融合：Anything占比重0.1与X1融合得到模型A1，重复上述操作两次得到A3；OrangeMix使用同样的方法进行融合得到B3。最后A3模型和B3模型对半融合。（不要问，问就是玄学）The V2 version is a fusion model. Because the trained model is over-fitting, model fusion is used to improve generalization.A model X is obtained by training, and then two copies (recorded as X1 and X2) are fused with the AnythingV3 and OrangeMix models respectively: the proportion of Anything is 0.1 and X1 is fused to get model A1, and the above operations are repeated twice to get A3; OrangeMix uses the same method to fuse to get B3. Finally, the A3 model and B3 model are fused in half. (Don't ask, just because it's metaphysics)

Files with SHA256: 16ebc12220e4b5e831e95ce598293feb9a7c8758c1df7e305a8740c098b75844