<h1 id="intro.-y75qe5fb2">Intro. 简介：</h1>A style LoCon trained on pony-based model images collected from Civitai site with "most collections" and "most reactions".这是一个训练自Civitai上点赞最多和收藏最多的pony系模型图片的画风LoCon。This lora does not intend to simulate any specific artist style or technique. It MIGHT reflects community taste and the visual attractiveness of a picture to a certain extent. Styles may change subtly depending on different prompts.这个lora并不意于还原某个特定的画师画风或者绘画技巧。它在某种程度上可能反应了社区审美和图片的视觉吸引力。 不同的提示词下可能会有微妙的画风变化。<h1 id="usage-8kxup8w9f">Usage 使用方法：</h1>Versions before V2 do not have specific trigger words. Please use the quality tags provided with the corresponding model.For V3 and later versions, the following tags were trained:V2以前的版本没有特定触发词。请使用对应模型自带的质量提示词。对于v3及后续版本，训练了以下标签：positive:<pre><code>masterpiece, best quality, very aesthetic</code></pre>negative:<pre><code>worst quality, low quality, displeasing</code></pre>你可以在此基础上编辑提示词。<h1 id="data-generation-hue452llf">Data Generation 数据版本：</h1><h2 id="v6:-je4q1aox1">v6:</h2>Added over 500 new images, some of which are selected from Flux. I removed some older images that I deemed to be of lower quality.The total number of images in data set now exceeds 3,000, with more than 20 concepts manually enhanced/edited across 6 versions of the dataset.The model’s rank has been increased as well.新添加了500+张图片其中有一部分选自flux生成的图像。删除了一部分我认为品质不佳的旧图片。现在总的图片数超过了3000，6个版本的数据总共手动增强/修正了20多条概念。增加了模型的rank。<h2 id="v5.9:-irqtg0puz">v5.9:</h2>The model's performance is not as expected, but I believe the images in the training data set are just fine. I'm planning to adjust the tags manually and see how will the results change.模型效果不如预期，但是我觉得训练集的图片本身应该没什么问题。打算先手动修正tag看看效果。2025/1/3更新：Manually updated some tags, but they seem unrelated to brightness and colors. Tentatively guessing it might be related to noise offset.手动更新了一部分标签，但是它们和明暗色彩无甚干系。暂时猜测可能和noise offset有关。<h2 id="v5:-qvonms90p">v5:</h2>The dataset has been expanded to 2,154 images, with around 1,000 Pony images as the primary training target.Although V-pred models can use LoRA trained on Eps-pred based models, the output quality drops significantly. This version will be trained separately on two different types of models.Recent versions of NoobAI exhibit noticeable artifacts, but the 'jpeg artifact' tag from Danbooru doesn’t seem to work effectively. To address this issue, about 30 typical and visually noticeable images were specifically selected as negative examples.An phenomenon has been observed: Pony v6 and NoobAI tend to generate a triangular lift at the roots of hairstyles with sidelocks. On Danbooru, this lift is sometimes tags as 'hair intakes' or 'curtained hair,' but Pony applies this structure to every character. This is a key reason why hairstyles generated by Pony often don't match the intended design during character training. A similar issue was observed with NoobAI. My guess is that this feature is prevalent in a dataset outside of Danbooru and was not correctly tagged.The images in the dataset were filtered, and about two-thirds were correctly annotated. Currently, adding 'hair intakes' to the prompt might somewhat alleviate this issue, but I haven’t found a complete fix for it yet.数据集扩充到2154张图。其中作为主要训练目标的pony图片约1000张。虽然V-pred模型也能使用基于Eps-pred技术的模型训练的lora，但是生成质量会大打折扣。这个版本将会分别在两个不同类型的模型上训练。noobAI近期版本有比较明显的伪影，但是danbooru上的“jpeg artifact”并没有起作用。因此专门针对这个问题选择了约30张较为典型的、肉眼可见的图片作为负面案例。观察到一个现象：pony v6和noobAI在生成有侧发的发型时，倾向于在发根处生成一个三角形的翘起。在danbooru里，这种翘起有时会被标注为“hair intakes”和“curtained hair”，但是pony会给每一个角色都套上这样的结构。这也是pony训练角色时，发型训练不像的一个重要原因。noob也观察到了类似的现象，我的猜测是danbooru以外的某个训练集大量存在这个特征，但没有对这个特征进行正确标注。对数据集里的图片进行了筛选，其中约2/3的图片进行了正确的标注。现在，在prompt里写上“hair intakes”可能可以一定程度上减轻这个现象，但是我还没有找到根治这个毛病的办法。<h2 id="v4:-asbjwur3n">v4:</h2>Partially optimized the dataset tags. Trained based on <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1022833">NoobAI Epsilon-pred v1</a> .Pony-based models have a strong tendency to generate earrings, ear piercing, and other types of accessories, sometimes messing up the ear structure of characters. I reorganized the related tags, cropped and manually edited some images in the dataset with minor structural issues, and removed pics that were too difficult to fix.对数据集的标注方式进行了部分优化。基于NoobAI Epsilon-pred v1训练。Pony系模型有很强烈的生成耳环、耳钉以其他类型的耳部饰品的倾向，有时还会破坏人物耳部的结构。对相关的标注进行了整理。剪裁、手工修改了数据集中一部分结构错误不严重的图，剔除了一些太难修改的图片。<h2 id="v3:-v5pauzm92">v3:</h2>Dataset extended to 1429 images, including examples with positive tags and negative tags.774 of the images are the most "wanted" style.Trained on <a target="_blank" rel="ugc" href="https://civarchive.com/models/795765?modelVersionId=889818">Illustrious v0.1</a>.数据集扩展到了1429张图片，包括了正反两种例子。其中774张是训练的目标风格。基于Illustrious v0.1训练。<h2 id="v2:-l3tov2gru">v2:</h2>Dataset extended to 374 images. Use quality tags and aesthetic tags which comes with models to control generation quality.训练数据集扩展到了374张。尝试使用模型自带的质量提示词来稳定生成质量。<h2 id="v1:-sxhcgv4c6">v1:</h2>Trained 224 images from Civitai, 393 images for regularization.Trained 2 versions based on <a target="_blank" rel="ugc" href="https://civarchive.com/models/260267?modelVersionId=403131">Animagine v3.1</a> and <a target="_blank" rel="ugc" href="https://civarchive.com/models/257749/pony-diffusion-v6-xl">Pony v6</a>.训练了C站上224张图片，393张正则数据集。有Animagine v3.1和Pony v6两个版本。<h3 id="test-ver.4:-gn79wmymh">test ver.4:</h3>It is a little bit underfitted but still works. I found that those quality tags and authentic tags (best quality, masterpiece, very aesthetic, ...) Animagine v3.1 has been trained can change the art style generated by this checkpoint. Fixing it in the next test version.有些欠拟合但是目前是有效的。我发现Animagine v3.1自带的质量控制词和美学提示词会改变生成图片的画风，所以这个实验版本需要不填写质量词。下一版会修复。<h3 id="-q7lx3mz73"></h3>

Pony: People's Works v1-v6

Base model: Animagine v3.1batch size=12dimension=32 conv_dim=8alpha=16 conv_alpha=2learning rate: U-net lr=6e-4 Te lr=2e-4optimizer: AdamW训练策略：训练数据集：374张Cvitai上的pony v6图片，393张正则图片训练总共21个epoch，每个epoch中重复5次训练图片、2次正则图片使用了 cosine with restarts 调度器，第一个epoch（323步）作为预热；剩余的20个epoch分为5个循环，每个循环中有4个epoch（也就是说每个cycle里训练了20次训练图片）每张图片训练了5+5×4×5=105次。Training strategy: Training dataset: 374 selected pony v6 pics from Civitai, 393 regularization imagesTrained 21 epochs, each epoch repeated training training data 5 times and reg data 2 times.Applied cosine with restarts scheduler, first 1 epoch (323 steps) for warmup and divided rest 20 epochs into 5 cycles, each cycle contained 4 epochs (training images were trained 20 times per cycle).Trained 5+5×4×5=105 times for one pic totally.

Files with SHA256: 5a81485da9668a285f9e65c771f4099ba580ff93cf4913a69a35c66333313af7