Be sure to check the version of LORA.Awaken_DuBusi_NoobAI was trained using base model NoobAI-epsilon-pred-1.1version.Note that lora trained on e-pred version will have unexpected effect on v-pred version.Awaken_DuBusi_veritas was trained using base model homoveritasXL_v10NAIXLEPSV11 by Kairen92.For Awaken_DuBusi_NoobAI, there are 2 versions, the one in zip file(<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:14px">703.9 MB) was trained at lower steps so it gives more anime style. The first suit picture was generated using this LORA.Use facedetailer in comfyui or adetailer in webui or inpaint to generate more detailed eyes.If excessive fingers/toes, try using prompt "4fingers"/"4toes".For Awaken_DuBusi_veritas:All pics are generated using HighresFix, but resolution higher than 1024*1024 using euler_a could possibly generate decent enough primitive pic.Full body pics are generated with face inpainting.Suggested LORA weight : 0.65 or 0.6~0.8. Lower LORA weight has more generalization capability and prompt adherence.Suggested sampler : euler_a or dpmpp_2m+karras. When using dpmpp_2m+karras denoise better be not less than 0.5 otherwise generated pics won't be fullly denoised .Upscale + HighresFix is required to provide better performance. Upscale image by 1.2~2(depending on your initial resolution) then I2I( or hires-fix) at denoise rate around 0.4 using euler_a. Alternatively, you can also upscale image only slightly , for instance, by 1.2, in this case, a higher denoise rate like 0.7 is allowed, which means dpmpp_2m+karras is allowed as well.Personally, I prefer euler_a, but there are times when the other works better, so you should try it yourself to see which yields the best performance for you.If you are experiencing extra fingers or what have you, try lower lora weight, denoise rate or upscale rate.If you wish to generate a full body pic with detailed face, detailed eyes or face more resembling original character, after HighresFix, face inpainting or Detailer is requried.Note that upscale is also needed before face inpainting and the lower percentage of the image occupied by the face, the more upscale rate it requires to provide more detailed and resembling face.For instance, if the primitive image is 768*1024, the character's face is most likely to be broken, u need to upscale image by 2 then highres-fix this image. Furthermore, if it's full body or the face is still borken or does not resemble the character, u need to once again upscale the image by 2 ,so it's x4 the initial resolution in all.Additional tips:1. When face inpainting, you can use prompts like "close-up, portrait" then delete "full body" and any other prompts that are not relative to face, this will enable the model to use close-up or portrait training data with better face and eye detailes to generate face. If the mask covers the face only, a higher rate of denoise like 0.5~0.6 and sampler method dpmpp_2m + karras are allowed. Conversely, if you do not wish to change prompts or the mask covers background and any other objects , use euler_a and lower denoise rate like 0.3, but the performance might be diminished .For example, the face in "sitting on throne" pic was inpainted in comfyui and the mask covers only the face including ear and part of the beard and hair but not the braid or background. Beofre inpainitng, the image was upsacled by 2. The sampler method is dpmpp_2m+karras ,denoise rate is at 0.6.2. You can use prompt "glowing eyes" to make the eyes glow which resembles the original character's eyes, but the ending result may be disastrous without high resolution+face inpainting. Conversely using "glowing eyes" as a negative prompt could make eyes less likely to be broken.底模使用的是<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">homoveritasXL_v10NAIXLEPSV11，由 Kairen92制作。所有图片均使用HighresFix生成，但是初始图片如果分辨率大于1024*1024的情况下，使用euler_a也可能产出效果还可以的图片。全身像通过面部局部重绘技术生成。建议参数配置：LORA权重：0.65 或 0.6~0.8（较低权重可增强泛化能力与提示词遵循度）采样器：euler_a 或 dpmpp_2m+karras（后者去噪强度建议≥0.5，否则图片可能残留噪点）核心流程：如果像效果更好一些，需要执行放大+HighresFix：先将图像放大1.5~2倍，再以euler_a采样器进行图生图（或hires-fix），去噪强度约0.4，当然，你也可以尝试放大较小的倍数，比如1.2倍，此时可以使用更高的去噪强度，比如0.7，这也意味着dpmpp_2m+karras也可以用。个人觉得前者综合更好，但有时候后者效果有时候反而会好一点，但是到底什么组合效果最好，还是需要自行尝试。全身像面部优化：生成全身像后，需通过面部修复或Detailer工具增强面部细节/眼睛/角色相似度修复前放大原则：面部占图像比例越小，所需超分倍率越高（例：原始图768*1024时，建议先2倍超分→高清修复；若面部仍破碎/不似角色，需再2倍超分进行局部修复→总计4倍分辨率）问题排查：出现多余手指等问题 → 尝试降低LORA权重及去噪强度，或者降低放大倍率。面部修复进阶技巧：提示词优化：删除"全身"等无关标签，添加"特写,肖像"类提示词，可激活模型的面部细节训练数据若仅遮盖面部区域（含耳朵/部分须发），允许：去噪强度0.5~0.6，也可以使用 dpmpp_2m+karras采样器若遮盖背景/其他物体，或者不想修改提示词，建议：euler_a采样器 + 去噪强度0.3（效果可能受限）发光眼睛实现：使用"glowing eyes"提示词可模拟角色原版眼睛⚠️ 未配合高分辨率+面部修复时易产生崩坏相反，使用"glowing eyes"作为负面提示词，可能让眼睛不容易崩操作实例："王座坐姿"图的面部修复流程：①2倍超分原始图② 在ComfyUI中仅遮盖面部（含耳朵/须发根部，排除辫子与背景）③ 增加"close-up, portrait"等面部特写提示词，删除"full body"等提示词，尽可能只保留面部相关提示词④ 采用dpmpp_2m+karras采样器 + 去噪强度0.6生成

Awaken DuBusi - Soul Land II

This LORA was trained using base model NoobAI.If excessive fingers/toes, use prompt "4fingers"/"4toes".use facedetailer in comfyui or adetailer in webui or inpaint to generate more detailed face.

Files with SHA256: 7586e8f45289c40b1cbaf5b26b1b0d4591c862179d60c984995746dcd4ac7f67