模型介绍: 使用真人大模型进行的训练,推荐大模型使用麦橘系列,aw系列,墨幽等真实模型,不建议在比较原始的模型比如chillout和basil等模型上使用,使用8400VAE。数据集为450张真人模特的图片,可以直出网红模特类型的美女不需要脸部或者高清修复,推荐使用25步以上的DPM++系列,cfg6-8,分辨率512x768出图,推荐lora使用在0.4-0.8范围内,推荐使用质量词语best quality,masterpiece,(photorealistic:1.4),raw photo,(Authentic skin texture_1.3。可以不用人脸修复插件即可达到很优秀的面部效果,因为使用的主要为半身肖像训练集,建议使用half body之类的半身或者近距离的词汇,防止脸部崩坏。 For training using real person models, it is recommended to use real models such as the magicmix series, aw series, and Mo You series for large models. It is not recommended to use these models on more primitive models such as Chilout and Basil, and instead use 8400VAE. The dataset consists of 450 images of real-life models, which can directly depict the beauty of the internet celebrity model type without the need for facial or high-definition repair. It is recommended to use the DPM++series with more than 25 steps, cfg6-8, resolution of 512x768, Lora is recommended to be used in the range of 0.4 to 0.8, and the quality words best quality, masterpiece, (photosensitivity: 1.4), raw photo are recommended, Authentic skin texture_1.3. Excellent facial effects can be achieved without the use of facial repair plugins, as the main training set for half body portraits is used. It is recommended to use half body or close range vocabulary to prevent facial collapse. ~~~~~~~~~~~~~~~~~~~~~~~o(* ̄︶ ̄*)o~~~~~~~~~~~~~~~~~~~~~~~ 示例提示词和参数: Prompt: best quality,masterpiece,(photorealistic:1.4),raw photo,(Authentic skin texture_1.3),1girl,solo,bangs,long hair, Negative prompt: (worst quality, low quality,NSFW:2),EasyNegative,ng_deepnegative_v1_75t, Steps: 30, CFG scale: 7, Sampler: DPM++ SDE Karras, Seed: 2509588528, Size: 512x768, Model: majicmixRealistic_betterV2V25, Version: v1.5.1, Model hash: d7e2ac2f4a, "EasyNegative: c74b4e810b03, AddNet Enabled: True, AddNet Model 1: reallora-000004(a8aa3ffcffa4), AddNet Module 1: LoRA, AddNet Weight A 1: 0.7, AddNet Weight B 1: 0.7, ng_deepnegative_v1_75t: 54e7e4826d53" |
2人表示很赞