<h2 id="ck-rf0.3-v1">关于CK RF0.3 V1</h2>基于ChenkinNoob-XL Rectified-Flow 0.3版本（<a target="_blank" rel="ugc" href="https://civarchive.com/models/2363696?modelVersionId=2753273">https://civarchive.com/models/2363696?modelVersionId=2753273</a>），使用LoRA_Easy_Training_Scripts训练，仅训练UNET，参数见About this version。与其他版本的主要区别在于训练40epoch以后的质感和构图增强，以及关闭Text Encoder训练后的文字错乱减少。此外触发词也有修改。<h3 id="regarding-ck-rf0.3-v1">Regarding CK RF0.3 V1</h3>Based on the ChenkinNoob-XL Rectified-Flow 0.3 version (<a target="_blank" rel="ugc" href="https://civarchive.com/models/2363696?modelVersionId=2753273">https://civarchive.com/models/2363696?modelVersionId=2753273</a>), trained using LoRA_Easy_Training_Scripts with only the UNET component trained (parameters detailed in "About this version"). The primary differences from other versions include enhanced texture and composition after 40 epochs, reduced text garbling by disabling Text Encoder training, and modifications to trigger words.<h2 id="noob-v-pred-3.0">关于NOOB V-PRED 3.0</h2>使用Kohya SS GUI基于NOOB V-PRED V1.0训练。修改训练参数，增加了训练轮数，降低每轮重复次数防止过拟合，现在Lora对于细节理解更好。具体训练细节见右侧版本相关。<h2 id="regarding-noob-v-pred-3.0">Regarding NOOB V-PRED 3.0</h2>Training based on NOOB V-PRED V1.0 using Kohya SS GUI.Adjusted the training parameters by increasing the number of training epochs and reducing the repetition frequency per epoch to prevent overfitting, resulting in better detail comprehension by LoRA.For specific training details, refer to the related version on the right<h2 id="noob-v-pred-2.0">关于NOOB V-PRED 2.0</h2>使用OneTrainer基于NOOB V-PRED V1.0训练。筛选训练集，增加了训练次数，来获得更好的色彩和效果。删除了style-abstract，因为NOOB对于这类艺术风格的理解很差。具体训练细节见右侧版本相关。<h2 id="about-noob-v-pred-2.0">About NOOB V-PRED 2.0</h2>Train using OneTrainer based on NOOB V-BED V1.0.Filter the training set and increase the number of training repeats to achieve better colors and effects.The triggerword s-o (style-abstract) has been removed because NOOB has a poor understanding of this type of art style.Please refer to the relevant version on the right for specific training details.<h2 id="noob-v-pred-v1.0">关于NOOB V-PRED V1.0</h2>使用NOOB V-PRED V1.0训练。减少了训练数据来获得更好的效果。现在它在monochrome方面表现更好。细分了触发词来生成不同类型的艺术风格，s-a的意思是style-abstract，s-o的意思是style-oil painting，s-m的意思是style-monochrome。具体训练细节见右侧版本相关。<h2 id="about-noob-v-pred-v1.0">About NOOB V-PRED V1.0</h2>Trained based on NOOB V-PRED V1.0.I reduced the training data for better effect. Now it performs better in monochrome.I divided the trigger words to generate different types of art styles, where s-a means style-abstract, s-o means style-oil painting, and s-m means style-mornochrome.It is recommended to use NOOB V-PRED 0.5 as the generative model for better aesthetic performance.Please refer to the about version on the right for specific training details.<h2 id="noob-e-pred-v1.0">关于NOOB E-PRED V1.0</h2>使用NOOB E-PRED V1.0训练。重新搜集了训练数据，细分了触发词来生成不同类型的艺术风格，s-a的意思是style-abstract，s-o的意思是style-oil painting，s-m的意思是style-monochrome。建议使用NOOB V-PRED 0.5作为生成模型，有更好的美学表现。具体训练细节见右侧版本相关。<h2 id="about-noob-e-pred-v1.0">About NOOB E-PRED V1.0</h2>Trained based on NOOB E-PRED V1.0.I recollected the training data and divide the trigger words to generate different types of art styles, where s-a means style-abstract, s-o means style-oil painting, and s-m means style-mornochrome.It is recommended to use NOOB V-PRED 0.5 as the generative model for better aesthetic performance.Please refer to the about version on the right for specific training details.<h3 id="ponyxl-experimental">关于PonyXL-experimental</h3>尝试以PonyXL作为底模进行训练，得到的结果勉强靠近了3.0，但是在影像细节上缺少很多质感，目前仍在探索更好的版本。训练参数见版本相关。<h3 id="about-the-ponyxl-experimental">About the PonyXL experimental</h3>Attempting to train with PonyXL as the base model resulted in a result that was barely close to 3.0, but lacked a lot of texture in image details. We are still exploring a better version. See about the version for training parameters.<h3 id="3.0">关于3.0</h3>在3.0版本中，使用了多分辨率噪声技术（金字塔噪声），没有使用正则化，所有图像都被应用到训练中。光影效果似乎变强了，另外对于规则物体的绘画也更灵敏。缺点是似乎更喜欢画建筑物，在画人物时需要增加1girl或1male的权重。训练参数如下：multires_noise_iterations="6"multires_noise_discount=0.3In 3.0 version, mutlires noise was used. No Regularization. All images were used in training.The light and shadow seem to have become stronger, and the painting of regular objects is also more sensitive. The disadvantage is that it seems to prefer to paint buildings, and when drawing characters, you need to increase the weight of 1 girl or 1 male.Here's the parameter of this version:multires_noise_iterations="6"multires_noise_discount=0.3<h3>导言</h3>这是一个用于绘制苏联海报风格的LyCORIS。当权重较低时，它偏向模型本身的风格，而权重较高时则会更偏向海报风。目前仅在二次元模型上测试过，主要测试模型为AOM3A，viewerMix，使用效果比较好，如果在偏真人的模型上可能可以得到更偏向于原版海报的效果。<h3>训练参数</h3>以NAI Final作为底模训练，使用了36张图片作为训练集，35张图片作为回归训练集，在训练前进行了镜像增强，每张图片重复训练10次，训练10个epoch，最终步数为7000步左右，以下为主要设置参数：network_dim=32network_alpha=32keep_tokens=4conv_dim=4conv_alpha=4lr="1e-4"unet_lr="1e-4"text_encoder_lr="1e-5"batch_size = 2<h3 id="introduction">Introduction</h3>This is a LyCORIS of CCCP poster styles. When the weight is low, it tends to favor the style of the model itself, while when the weight is high, it's more like the poster style.At present, it has only been tested on the anime model. The main test models are AOM3A and viewerMix, and the effect is relatively good. The effect of this LyCORIS may be more favorable to the original poster on the realistic model.<h3 id="training-parameters">Training parameters</h3>It is trained based on NAI final. I used 36 images as a training set and 35 images as a regression training set, image enhancement was performed before training. Each image was repeatedly trained 10 times for 10 epochs, with a final number of steps of about 7000. The following are the main setting parameters:network_ dim=32network_ alpha=32keep_ tokens=4conv_ dim=4conv_ alpha=4lr="1e-4"unet_ lr="1e-4"text_ encoder_ lr="1e-5"batch_ size = 2

CCCP poster style

在2.0版本中，增加了训练集的数量，一共使用了264张图片进行训练，在训练前进行了镜像增强，半数图片作为正则化图集使用，每张图片训练3次，并启用了offsetnoise，设置值为0.1。2.0版本中的光影质量更好，并且该黑的地方能黑下来。除了使用Viewermix进行测试外，还使用了三馅混沌<a target="_blank" rel="ugc" href="https://civitai.com/models/20330/three-delicacy-wonton-mix">Three Delicacy Wonton (三餡馄饨Mix) | Stable Diffusion Checkpoint | Civitai</a>模型进行了测试，效果不错。In version 2.0, the number of training sets has been increased. A total of 264 images have been used for training, and image enhancement has been performed before training. Half of the images have been used as regularized atlases. Each image has been trained three times, and offsetnoise has been enabled, with a setting value of 0.1.The light and shadow quality in version 2.0 is better, and the dark areas can be darkened.In addition to using the Viewermix for testing, we also used the three stuffing chaos Three Delicacy Wonton | Stable Diffusion Checkpoint | Civitai model for testing, and the results were good.

Files with SHA256: 2b1147b00d27098fb427bbcd76e8374f70a12a3678fe06cd42d02a41987abb7e