<ul><li>一种“较”真实系二次元风格，在经过一些神秘的lora配比后获得的风格（大嘘），利用这个奇妙风格后生成的图片在魔塔上进行训练</li><li>krea2版测试时使用的底膜为官方krea2 turbo int8，其它模型不知道能不能适配</li></ul><ul><li>A realistic anime style, obtained after some mysterious lora matching, and the images generated using this wonderful style are trained on the ModelScope Civison.</li><li>The base model used during testing is the official krea2 turbo int8. I'm not sure if other models can be adapted.</li></ul><h3 id="anima-v1.0">Anima v1.0</h3><ul><li>有了较为明显的发丝和衣服褶皱，但还远远还不够</li><li>后续还需要进行服装、头发细节等特化训练，可能需要大量的时间</li><li>强烈建议加入触发词real skin, real cloth, soft natural lighting, detailed skin, fabric folds, hair strand, wrinkled fabric</li></ul><ul><li>There are relatively noticeable hair strands and clothes wrinkles, but it's far from enough</li><li>Subsequent specialized training for clothing and hair details is also required, which may take a significant amount of time</li><li>It is strongly recommended to incorporate trigger words "real skin, real cloth, soft natural lighting, detailed skin, fabric folds, hair strand, wrinkled fabric"</li></ul><h3 id="anima-v0.9">Anima v0.9</h3><ul><li>v0.9为预览版，效果不佳，作为基准画风为后续改进提供参考</li><li>在anima-base上进行训练，使用其他模型作为底膜时不知道能不能获得正确的效果</li><li>后续会对数据集进行改进以达到正确的风格效果</li></ul><ul><li>v0.9 is a preview version with poor effects, serving as a benchmark style for future improvements</li><li>When training on anima-base and using other models as the base layer, it is uncertain whether the correct effect can be achieved</li><li>The dataset will be improved subsequently to achieve the correct stylistic effect</li></ul><h2 id="v1.2">v1.2</h2><ul><li>在优化了数据集和打标后，质感获得了进一步提升，但仍有部分模糊出现以及块状感</li><li>推荐搭配Detailer - beauty for Krea2</li></ul><ul><li>After optimizing the dataset and labeling, the texture quality has been further improved, but there are still some blurring and blocky effects</li><li>Recommended combination: Detailer - beauty for Krea2</li></ul><h2 id="v1.1">v1.1</h2><ul><li>优化了一下数据集，再进行更高一点的分辨率训练获得了1.1版本（</li><li>目前二采基本不会有太大的“硬化”效果，但可能会出现一些奇怪的细节和细节错乱</li><li>无论一采还是二采，在衣服褶皱细节方面已经接近原数据集80%的质感（感觉、应该、大概），但是头发依旧不太行，会有种3d方块的感觉，跟数据集中的柔和发丝还是差别很大。不过远距离视角下的人物头发确实是柔和的质感，近距离视角真跟方块一样还会部分细节错乱（不懂是不是过拟合）</li><li>目前正在尝试在现有lora上进行加训，如果效果还是不好会尝试再搞一份数据集重练，再不好就只能换训练网站或云端训练了<s>（有一说一魔塔的训练参数少的可怜，打标也不能自己打，感觉目前问题大都出现在打标上）</s></li><li>推荐0.8的强度再加个1强度的Detailer - beauty for Krea2</li><li>后续有时间会考虑练个anima版本</li></ul><ul><li>After optimizing the dataset and conducting training at a higher resolution, we obtained version 1.1</li><li>Currently, secondary sampling generally does not have a significant "hardening" effect, but it may introduce some strange details and disruptions in details</li><li>The texture of the hair is still not very good, giving a 3D blocky feel, which is quite different from the soft strands in the dataset. However, from a distant perspective, the character's hair does have a soft texture, but up close, it really looks like blocks with some details scrambled</li><li>Currently, we are attempting to conduct additional training on the existing LoRa platform. If the results remain unsatisfactory, we will attempt to create a new dataset for retraining. If that fails, we will have to switch to a different training website or cloud-based training</li><li>Recommend a strength of 0.8 and add a 1-strength Detailer - beauty for Krea2</li><li>I'll consider practicing the anima version when I have time later</li></ul><h2 id="v1.0">v1.0</h2><ul><li>经测试后效果不是特别好，并且不适用于潜空间放大二采，即使是一采生成出来的图也有一些“硬化”的质感。当然，上述问题可能在后续的版本中进行改进</li><li>此风格可能会导致画面有种灰暗感</li><li>推荐增加“Anime in real”和“Detailer - beauty for Krea2”来获得更强的质感</li></ul><ul><li>After testing, the effect is not particularly good, and it is not suitable for secondary mining of latent space amplification. Even the images generated from primary mining have a somewhat "hardened" texture. Of course, the aforementioned issues may be improved in subsequent versions,perhaps.</li><li>This style may give the image a gloomy feel.</li><li>It is recommended to add "Anime in real" and "Detailer - beauty for Krea2" to achieve a stronger sense of texture.</li></ul>

真实细节二次元——Realistic details anime

rda v1.2.safetensors

Files with SHA256: 6065a85a941846abc3469c7377ebca439b465fca27f50be808035fd962635e5b