<h2 id="-kj581uuw1">非常希望大家能将生成的图片上传到这里！充足的反馈对于改进模型来说是很有必要的！！</h2><h2 id="hey-guys!-really-hope-you-can-upload-the-pics-you-make-here!-getting-lots-of-feedback-is-super-important-for-making-the-model-better!!-vvs2nrrw7">Hey guys! Really hope you can upload the pics you make here! Getting lots of feedback is super important for making the model better!!</h2><h1 id="elyponyv15-4efbv2f5s">ElyPonyV15</h1>ElyPonyV15 Guide:1. I created approximately 300 high-quality images using AnythingXL, based on my personal preferences. I then utilized Claude 3.5 Sonnet to tag these images using natural language. Subsequently, I trained a LoRA model on a combination of T-ponynai3 and PonyV6. Finally, I merged this LoRA into T-ponynai3, resulting in ElyPonyV15. This version differs slightly from ElyPonyV10, featuring cooler colors, softer brush strokes, and slightly paler shades.1.我根据我自己的风格偏好，使用AnythingXL生成了接近300张的高质量图片作为训练集，并用Claude3.5 sonnet以自然语言对训练集中的图片进行了标注，之后在T-ponynai3和PonyV6的融合模型上训练了LORA，然后将LORA融合进了T-ponynai3进行微调，得到了ElyPonyV15。ElyPonyV15的风格和ElyPonyV10略有不同，相比ElyPonyV10，ElyPonyV15更偏向冷色调，笔触更柔和，颜色也变得稍微寡淡了些。2. I used Claude3.5 sonnet to tag the training pics in natural language. This helps it work better with natural-language prompts. Here is the example:2.我使用Claude3.5 sonnet对训练集中的图片用自然语言进行了标注，以提高其在自然语言prompts下的性能，示例如下：<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ff33059b-d218-4f5d-bd0a-e55dcc04c18e/width=525/ff33059b-d218-4f5d-bd0a-e55dcc04c18e.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9e88a55a-8bc0-4a56-90d5-0249087f15b7/width=525/9e88a55a-8bc0-4a56-90d5-0249087f15b7.jpeg" />You may get better results with this kind of prompts使用这种风格的提示词或许会获得更好的结果3. For the special prompts-"score_x_up" in the Pony series, there are two sets of templates you could choose. The first one is what Pony's creator recommends-"score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up". This set will give you sharper lines, but the colors might look a bit washed out. The second set is what most people usually go for: "score_9, score_8_up, score_7_up". With this one, you'll get nicer colors, but the lines won't be as crisp. Just pick whichever you like best. (Personally, I lean towards the first set of templates.)3.针对pony系列的专有prompts-"score_x_up"，我建议使用如下两套模板。第一套就是Pony的作者所推荐的"score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up"，这套prompts模板可以带来更锐利的线条，但也会让色彩偏淡。第二套就是大部分人常用的"score_9, score_8_up, score_7_up"，这套prompts模板能带来更好的色彩，但是也会让线条变得不那么锐利。选择你们喜欢的就好。（我个人更倾向于第一套模板）4.Same as ElyPonyV10, ElyPonyV15 needs prompt "PEX_style" to keep the art style steady. 4.和ElyPonyV10一样，你需要添加"PEX_style"以稳定画风5.Adding "rating_explicit" can really boost the chances of getting good hands and good anatomy, and it works way better than any embeddings. But it'll make the pics more, uh...spicy, even if you don't use any NSFW prompts.5.添加rating_explicit可以大幅增加手部正常以及肢体协调的概率，比任何embeddings都有用，不过会让生成的图片更涩，即使你没有使用任何NSFW的prompt（只有画涩涩的时候小马才能全力发功，真够黑色幽默的哈哈）<h1 id="elyponyv10-i5fcfgcda">ElyPonyV10</h1>模型介绍：ElyPonyV10可以算是一个小尝试，目的是让Pony稳定产出线条清晰、细节丰富、色彩鲜艳且画面干净整洁的图像，为此使用Anything XL获得了一些图片对ponyV6进行了finetune，但结果不是很理想，原因可能是训练集太少，因此我换了一种思路，我首先将T-ponynai3和ponyV6进行了融合得到了底模，目的是使用微调模型来稳定ponyV6的画风，让之后LORA融合进底模的成功率更高，之后选取了数张高质量的图片在T-ponynai3中使用差异法进行训练，部分训练集如下Model Introduction: ElyPonyV10 can be considered a small experiment. The goal is to enable Pony to consistently produce images with clear lines, rich details, and vibrant, clean colors. For this purpose, I got some images with AnthingXL and performed finetune on ponyV6, but the results were not very good, possibly due to a lack of sufficient training iamges. Therefore, I adopted a different approach. First, I merged T-ponynai3 and ponyV6 to create a base model with the aim of using a finetune model to stabilize the art style of ponyV6. Subsequently, I selected some high-quality images and trained the LORA in T-ponynai3 with the training method-"コピー機LoRA". Some of the training images are as follows<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ff5a78a0-4f6f-465e-9d37-fd4290a3d66b/width=525/ff5a78a0-4f6f-465e-9d37-fd4290a3d66b.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/842220b7-61bf-4b0b-9076-a0426526f11a/width=525/842220b7-61bf-4b0b-9076-a0426526f11a.jpeg" />最后将训练好的画风LORA融合进了T-ponynai3和ponyV6的混合模型中，得到了ElyPonyV10，下一步的目标是获取足够多的训练素材对ponyV6直接进行微调，以得到更接近训练素材质量的模型。Finally, I merged the LORA into the previously mentioned merged model of T-ponynai3 and ponyV6, resulting in ElyPonyV10. My next goal is to acquire sufficient training images to directly finetune ponyV6, aiming to get a model that is closer in quality to the training images.<h3 id="elyponyv10-7aerphqdd">ElyPonyV10使用说明：</h3>1.DPM++ 3M SDE貌似有更好的锐度和色彩表现，但其他采样器也不错，选择你喜欢的就好2.因为该模型是“LORA+底模”的构造，所以最好加上关键词“PEX_style”来稳定画风。3.如果想复刻示例图的画风，prompt可以直接照抄我的，最重要的是PonyV6的起手词不要缺，可以把score_9, score_8_up, score_7_up放到正面提示中，把score_6，score_5，score_4放到负面提示中。4.使用不同的prompt和negtive prompt会使画风产生改变，但基本风格应该不会变化太大，我没有实验太多。<h3 id=":-8lnrr27b6">下面是我常用的提示词:</h3>正面提示词：<code>((PEX_style)),score_9, score_8_up, score_7_up, best quality, high quality, masterpiece, Expressiveh, source_anime, masterpiece, japanese anime style, fantasy, (人物提示词), (背景提示词);</code>负面提示词:<code>score_6, score_5, score_4, 3d, watermark, signature, facial hair, beard, extra penis, extra head, missing eye, text, watermark, low-quality, signature, moir pattern, downsampling, aliasing, distorted, blurry, glossy, blur, jpeg artifacts, bara, compression artifacts, poorly drawn, low-resolution, bad, distortion, twisted, exaggerated pose, grainy, symmetrical, duplicate, error, pattern, beginner, pixelated, glitch, overexposed, bad-contrast, low quality, worst quality, black and white, black and white pencil sketch, pencil sketch, facial hair, hairy body, (fat, bulky), grab by himself</code><h3 id="elyponyv10-usage-instructions:-6uwgjv3ys">ElyPonyV10 Usage Instructions:</h3>1.you can get sharper images and more vibrant colors with DPM++ 3M SDE, but other samplers are good as well, just choose what you like2.Since this model is constructed using “LORA + base model”, it is best to add the keyword “PEX_style” to stabilize the art style.3.If you want to replicate the style of the example images, you can directly copy my prompts. It is important to include PonyV6's initial keywords, put score_9, score_8_up, score_7_up in Prompt and score_6, score_5, score_4 in Negative Prompt.4.Using different prompts and negative prompts may get different art styles, but the basic style won't vary too much. I haven't experimented extensively with this.<h3 id="here-is-the-prompts-and-negative-prompts-i-commonly-use-with-elyponyv10:-97xlnmfu9">Here is the prompts and negative prompts I commonly use with ElyPonyV10:</h3>Prompts:<code>((PEX_style)),score_9, score_8_up, score_7_up, best quality, high quality, masterpiece, Expressiveh, source_anime, masterpiece, japanese anime style, fantasy, (character prompts), (background prompts);</code>Negative Prompts:<code>score_6, score_5, score_4, 3d, watermark, signature, facial hair, beard, extra penis, extra head, missing eye, text, watermark, low-quality, signature, moir pattern, downsampling, aliasing, distorted, blurry, glossy, blur, jpeg artifacts, bara, compression artifacts, poorly drawn, low-resolution, bad, distortion, twisted, exaggerated pose, grainy, symmetrical, duplicate, error, pattern, beginner, pixelated, glitch, overexposed, bad-contrast, low quality, worst quality, black and white, black and white pencil sketch, pencil sketch, facial hair, hairy body, (fat, bulky), grab by himself</code>如果你在使用SDXL/PDXL系列的模型时经常出现图像崩坏，可以安装这个插件：<a target="_blank" rel="ugc" href="https://github.com/hako-mikan/sd-webui-prevent-artifact">https://github.com/hako-mikan/sd-webui-prevent-artifact</a>By the way, this plugin can prevent the "image corruption" of SDXL/PDXL series: <a target="_blank" rel="ugc" href="https://github.com/hako-mikan/sd-webui-prevent-artifact">https://github.com/hako-mikan/sd-webui-prevent-artifact</a>

Files with SHA256: c742022323bbe94a52922ffdc62f6f3ba9c5e308602f1b05ececc81f9abe129c