The LoRA is trained based on <code>Juggernaut_X_RunDiffusion.safetensors</code>：<a target="_blank" rel="ugc" href="https://civarchive.com/models/133005?modelVersionId=456194">Juggernaut XL - Jugg_X_by RunDiffusion | Stable Diffusion Checkpoint | Civitai</a><code>Saika</code> is the name of this girl, and also the trigger word of this lora. Add it in prompt, the generations will show her picture.<ul><li>想要出图很像Saika，则除了要带Saika之外，建议从`Saika, 1girl, solo, long hair, realistic, black hair, parted lip,`为出发点进行增删tag。(v2版本的LoRA会解决这个问题)</li></ul><ul><li>parted lips和smile这俩tag吸收了一部分Saika的面部特征，如果出图不像，建议加上parted lips或者smile</li></ul><ul><li>其他场景和服饰的描述不要太多，太多的话，会让人物不像Saika</li></ul><ul><li>不要直接出全身图，因为训练时候拟合Saika面部特征的人脸图片是512*512和1024*1024像素，全身图的832*1216，人脸占的区域估计是128*128或者更小，无法在这么小的区域里重现Saika的面部特征。</li></ul><ul><li>建议先出半身图或者大头贴，人物面部特征满意之后，用Outpainting扩图工具扩大到全身图。</li></ul><ul><li>不加入Saika的话，只用1boy，1girl都能正常出亚洲特征面孔，因此也是能泛化出优质asian boy和asian girl照片的合适lora。</li></ul><ul><li>由于没有512*512和1024*1024尺寸非常清晰的Saika肖像，所以实际1664*2432尺寸的半身图，人脸也并没有纤毫毕现的效果。后续如果有更多这个尺寸的清晰Saika肖像，会迭代出新的版本。</li></ul><ul><li>先batch count开个6或者9批量跑基础尺寸的图(batch size可以开2、4或更高，只要不cuda out of memory即可)，直到有结构比较好的图了(人脸不像也无所谓)。再锁定该种子，之后开hires。如果人脸不像，则图生图或者提高降低一下hires的denoising steps。</li></ul><ul><li>想出大头贴，用portrait。想出半身图，用upper body。这都是训练集里微调训练过的tag，用face、half body的话，没经过我微调，不知道它们对应什么图片特征。（全身图不建议出，上面提到过</li></ul><ul><li>如果面部扭曲、手指扭曲但身体结构很正常，那么seed不要改了，可以试试提高或者降低denosing steps。或者删除掉一些复杂的tag。</li><li>负面提示词务必加入blurry和watermark</li><li>训练集里有一些赵丽颖的古装照，导致赵丽颖的面部特征被映射到hanfu和chinese_clothes里了，如果出其他asian girls，面部总有些像赵丽颖，需要删掉hanfu、Chinese_clothes的tag，或者降低它们的权重。或者迭代步数降低到20以下。或者用kimono、Japanese_clothes，自动打标把一部分汉服打标为kimono了。用kimono也能触发汉服特征的服饰。或者加入lips，nose的描写，改变脸部五官的内容。</li></ul><ul><li>If you want to generate images that closely resemble Saika, it’s recommended to start with tags like <code>Saika, 1girl, solo, long hair, realistic, black hair, parted lips</code> and then add or remove tags accordingly. (This issue will be addressed in version 2.)</li><li>The tags <code>parted lips</code> and <code>smile</code> capture some of Saika's facial features. If the generated image doesn’t look like her, consider adding <code>parted lips</code> or <code>smile</code>.</li><li>Avoid including too many descriptions of scenes or outfits, as it can result in the character looking less like Saika.</li><li>Don’t generate full-body images directly, as the training focused on Saika’s facial features in images with resolutions of 512x512 and 1024x1024. In full-body images with 832x1216 resolution, the face region is only about 128x128 pixels or smaller, making it difficult to reproduce Saika’s facial features in such a small area.</li><li>It’s better to first generate half-body or close-up portraits. Once the facial features are satisfactory, use outpainting tools to expand the image to a full-body shot.</li><li>Even without including Saika, using only <code>1boy</code> or <code>1girl</code> can still produce faces with Asian features. This makes the LoRA suitable for generating high-quality images of Asian boys and girls.</li><li>Due to the lack of very clear 512x512 and 1024x1024 Saika portraits, half-body images at 1664x2432 resolution may not have ultra-detailed facial features. A new version will be released once more clear portraits of this resolution become available.</li><li>Start with a batch count of 6 or 9 at the base resolution (batch size can be 2, 4, or higher, as long as it doesn’t cause a CUDA out-of-memory error). Continue until you get a structurally sound image (even if the face doesn’t look perfect). Then lock the seed and use the hires function. If the face still doesn’t look right, consider regenerating or adjusting the denoising steps for the hires process.</li><li>Use <code>portrait</code> for close-ups and <code>upper body</code> for half-body shots. These tags were fine-tuned during training. Tags like <code>face</code> or <code>half body</code> weren’t fine-tuned by me, so their corresponding image features are unknown. (Generating full-body images is not recommended, as mentioned earlier.)</li><li>If the face or fingers are distorted but the body structure is fine, keep the seed and try adjusting the denoising steps up or down, or remove some of the more complex tags.</li><li>Make sure to include <code>blurry</code> and <code>watermark</code> as negative prompts.</li><li>The training set contains some traditional costume images of Zhao Liying, which led to her facial features being mapped to tags like <code>hanfu</code> and <code>Chinese_clothes</code>. If you’re generating other Asian girls and notice they resemble Zhao Liying, you can remove or reduce the weight of these tags. Alternatively, reduce the iteration steps below 20. You can also use <code>kimono</code> or <code>Japanese_clothes</code>, since some hanfu images were automatically labeled as kimonos during tagging. Using <code>kimono</code> may still trigger traditional Chinese clothing features. Additionally, adding descriptions like <code>lips</code> and <code>nose</code> can help alter the facial features.</li></ul>

Files with SHA256: 41c56580bb6a7809584971a7d987b410b96a14dce20f8c77bb25d168091cfe1b