<h1 id="20240914v3-owxsqv72q">20240914v3</h1>Naotora Ii from DEAD OR ALIVE 5Trigger Words: naotora iiThe training of this model and the images it generates are solely for learning purposes.Retrained with the accumulated experience, added a set of clothes. Modified some tags in the dataset. However, as the size of the dataset increases, maintaining it also becomes troublesome.<ul><li>The model's output may heavily rely on the checkpoint.</li><li>You could find the example prompts from the images above.</li><li>My prompts are basically composed in the order of [character traits] + [style] + [expression] + [clothing] + [camera and action] + [background], and you can delete or modify them as needed.</li><li>You could add '3D' in the negative prompt to reduce the model's 3d style.</li><li>Adding tag such as 'realistic', 'realism' could enhance the features of the character.</li><li>Recommended weight: 1~0.6, adjust as needed until the character's appearance meets your requirements.</li><li>Upscale needed for a better performance. Upscale value recommendation is around 1.3, denoising strength is 0.2</li><li>Facial distortion may easily occur in situations such as full-body shots. If there is facial distortion, consider using ADetailer for repair.</li></ul><h1 id="20240707-ew5bqgfc4">20240707</h1>This version has optimized the tags of the dataset, making it easier for this character to be reconstructed through the tags I provide.Currently, the images produced with "miko outfit" are prone to blurriness, which is likely due to the poor quality of the images in the dataset.<h1 id="20240526-x9t1wqluo">20240526</h1>The example tags could be found from the images above. Please remember to replace tag: &lt;lora:naotorali-000070&gt; or something like that on example with &lt;lora:naotorali&gt;The dataset were collected mainly from of screenshots of DOA5 and models created by users <a target="_blank" rel="ugc" href="https://www.deviantart.com/rolance">rolance </a>on DeviantArt.This time, around 140 images were collected, most of which are screenshots from various angles of the XPS models. The lora model trained in this way performs well.The main issue at present is that the tag annotations are not precise enough; sometimes the tags I provided cannot fully simulate several sets of costumes for this character. Additionally, during my training process, I found that I had made many mistakes in tagging. These issues may be fixed in the future.There are still a set of costumes for this character that can be added through further training, but that's a topic for later.I recall around 10 years ago, at that time, to create secondary works for these characters, there were mainly two ways: either you could draw, or you could find related models on internet for pose creation, and then import them into rendering software for rendering. Being able to produce a set of images in one or two weeks through this process was already considered a master. Now, most of the time on doing this kind of things is spent training models, and then you can generate dozens of sets of images in a short time, which is tens of times faster than before, and the quality of the images is almost identical to the original images. The greatest credit, in addition to technological progress, should go to the creators who share information online.

Files with SHA256: 1a4df0b80d225d23cd4b229b859af95517a9fc8a3a07def613b28a40e413383a