之前绫波丽的服装发白的原因是因为提示词中含有白色的描述，这次我使用特殊字符串代替了服装的描述，所以不会发白，只需要新的高噪声模型，低噪声模型用上个版本就可以了Previously, Rei Ayanami's clothing appeared washed out because the prompt contained descriptions of white. This time, I replaced the clothing description with a special string, so it won't appear washed out. Only a new high-noise model is needed—the low-noise model from the previous version works fine. Here is the prompt:绫波丽/Ayanami Rei informal dress：wearing se1fuku1t costume, hairstyle: ayanam1ha1r,绫波丽/Ayanami Rei Combat suit: wearing ayanam1su1t costume, hairstyle: ayanam1ha1r,明日香/asuka informal dress：wearing se1fuku1t costume, hairstyle: as5kaha1r. 明日香/asuka Combat suit：wearing as5kasu1t costume, hairstyle: as5kaha1r.如果大家有什么好的想法，或者想要训练的可以告诉我。这次我尝试了使用正则化训练，花费了我三倍的步数，万幸的是，的确解决了大部分问题，首先解决了单人战斗服在静止状态下出现的动漫脸和塑料皮肤，也解决了双人都穿着战斗服的视频几乎无可避免的塑料脸。现在只需要在双人视频开头加入"realistic，realistic woman，真实的皮肤质感"即可保证90%以上的情况都是真人脸，单人图不需要加入"真实的皮肤质感"，如果加入可能会导致发型出现一些改变，不过影响不大。同时我改变了数据集，让服装的材质变得更好，之前的服装看起来像一块步上面画了一些图案，现在他看起来是有光泽的橡胶质感，服装相似性也达到了90%以上，发型问题也得到了解决。不过白色胶衣的黑色部分似乎过度曝光严重，即使我的素材没有任何一张过度曝光的图片，我猜这可能是wan2.2自己的原因。为了避免这个问题，请加入对光线的描述。同样的，这次我使用了特殊字符串去标记服装，并在打标过程中不对服装进行描述，不过我仍然没有成功的将这些特殊的字符串和对应的服装联系起来，所以提示词仍然是我自己写的一大串文字，我仍然将他们放进了zip文件夹里，请仔细阅读。同时，我猜测这意味着你们可以将我的提示词直接翻译成英文使用。同样的这次的训练素材里也不包含任何未经审核的内容。如果你发现你的结果和我的相差过大，请确保你的wan2.2使用的是gguf q4及以上的量化的版本，不要使用fp8，虽然fp8看起来很大，但是效果甚至不如q4，同时不要开启加速。This time, I tried using regularization during training, which cost me three times the usual number of steps. Fortunately, it indeed resolved most of the issues. First, it fixed the anime-style faces and plastic-looking skin that previously appeared on single-character battle suits in static poses. It also solved the nearly unavoidable "plastic face" issue in videos where both characters wore battle suits.Now, for dual-character videos, simply adding prompts like "realistic, realistic woman, 真实的皮肤质感" at the beginning ensures over 90% of outputs feature photorealistic human faces. For single-character images, you don’t need to include "真实的皮肤质感"—in fact, adding it might slightly alter the hairstyle, though the impact is minimal.Additionally, I improved the dataset to enhance fabric quality. Previously, the outfits looked like flat cloth with printed patterns; now they have a glossy, rubber-like texture. Clothing fidelity has reached over 90%, and the hairstyle issues have also been resolved.However, the black parts of the white latex outfit appear severely overexposed—even though none of my training images are overexposed at all. I suspect this might be due to Wan 2.2 itself.To avoid this issue, please include a description of the lighting in your prompt.This time I used special strings to tag the outfits and avoided describing the clothing during the tagging process. However, I still haven’t successfully linked these special strings to their corresponding outfits. Therefore, I’m still relying on my own long, detailed prompt phrases, which I’ve again included in the ZIP folder—please read them carefully.I also suspect this means you can directly translate my prompts into English for use.As before, the training materials used in this LoRA contain no unreviewed or inappropriate content whatsoever.If you find your results differ significantly from mine, please make sure your Wan 2.2 is using a GGUF quantized version of q4 or higher. Do not use fp8—although fp8 may seem like a higher precision format, its performance is actually worse than q4. Also, do not enable acceleration.这是一个服装lora，也是一个cosplay lora。同时也是希望大家能给我一点建议这个lora只是一个尝试，因为我发现wan2.2训练动漫lora效果很差，所以改为训练cosplay lora我尝试在里面训练了三套衣服和两套发型，也就是绫波丽和明日香的校服和战斗服，校服的效果很好，战斗服也可以使用。另外可能是wan2.2自己训练集有太多的双马尾，我打标和wan2.2自带的双马尾重合了，所以明日香的发型不是那么好。将这个lora和任何其他真实风格的lora一起使用，效果会更好。单独使用这个lora有些时候会生成蜡像一样的画面，尤其是在双人场景，经常发生，我不知道具体原因，但是我的素材百分百都是使用真实画面的，所以如果有人知道这是为什么，请告诉我。我不知道为什么如果不使用<a target="_blank" rel="ugc" href="https://civarchive.com/models/1585622?modelVersionId=2261165">https://civarchive.com/models/1585622?modelVersionId=2261165</a>这个lora，就无法生成正确的效果，同时我也实在是不能理解wan2.2到底要怎么打标，我感觉我写的打标词全部失效了，wan2.2只会按照自己识别到的东西进行学习，所以我甚至要花一大截时间去琢磨怎么写提示词把我的素材里的东西复现出来，发这个lora出来也是希望大家能给我一点建议，能指导我关于如何打标。另一个问题是，你描述的动作越详细，效果会更好，如果不描述动作，可能会生成3d的画面，我并不能理解。也希望大家能给我一些建议。这个lora完全使用安全且健康的素材，不包含任何未经审查的内容。提示词很多很复杂，在zip文件里。This is a clothing LoRA, as well as a cosplay LoRA. I’m also hoping to receive some suggestions from the community.This LoRA is just an experimental attempt. I noticed that training anime-style LoRAs with Wan 2.2 yields very poor results, so I switched to training a cosplay-focused LoRA instead.I included three outfits and two hairstyles in the training: specifically, Asuka and Rei Ayanami’s school uniforms and battle suits. The school uniforms turned out very well, and the battle suits are usable too. However, Wan 2.2’s own training dataset seems to contain an overwhelming number of twin-tail hairstyles. Because my tags overlapped heavily with Wan 2.2’s built-in twin-tail representations, Asuka’s hairstyle didn’t come out as well as expected.When using this LoRA alone, it sometimes generates images that look like wax figures, especially in scenes with two people; this happens quite often. I don't know the exact reason for this, but since I use 100% real-life images as materials, if anyone knows why this is happening, please let me know.This LoRA works best when combined with other photorealistic (real-life style) LoRAs.I’ve also found that without using this specific LoRA—<a target="_blank" rel="ugc" href="https://civarchive.com/models/1585622?modelVersionId=2261165—the">https://civarchive.com/models/1585622?modelVersionId=2261165 </a>the correct results simply won’t appear. Moreover, I’m completely baffled by how tagging should be done for Wan 2.2. It feels like all the caption I carefully wrote are being ignored; the model seems to learn only based on what it itself recognizes in the images. As a result, I’ve had to spend a huge amount of time experimenting with prompts just to reproduce the elements present in my training data.One reason I’m releasing this LoRA is to ask the community for advice—especially guidance on proper tagging techniques for Wan 2.2.Again, the more detailed your action/pose descriptions, the better the output. Without them, you risk getting 3D-like renders, which confuses me greatly. Any suggestions would be greatly appreciated!All training materials used for this LoRA are safe, healthy, and contain no unreviewed or inappropriate content.The prompts are numerous and complex—they’re included in the ZIP file.

Files with SHA256: 542271fda5f7e9803a62693258fa0980aeab5cd728fb0efcf5862d68858037a3