<h3 id="heading-19">一、模型概述：</h3><ul><li>水墨风格lora模型，由将近150张水墨风的图片训练而成，</li><li>生成的图片带有中国水墨元素，水墨风的体现主要在背景和笔触方面，</li><li>优先测试了部分参数和加载大模型，更多的发现等待小伙伴们自己去挖掘......</li></ul><h3 id="heading-20">二、模型使用说明：</h3>1、推荐lora权重：0.6-0.9，添加触发词：shuimofeng，2、推荐加载基底大模型：<a target="_blank" rel="ugc" href="https://civarchive.com/models/6424?modelVersionId=11745">ChilloutMix</a>或者<a target="_blank" rel="ugc" href="https://civarchive.com/models/7371?modelVersionId=43978">ReV Animated_v1.2.2</a>，3、迭代步数（Steps）：25-40，CFG scale：10-15，Clip skip：2，4、推荐采样器：DPM++ SDE Karras（也可以试试其他采样方法），5、VAE： <a target="_blank" rel="ugc" href="https://huggingface.co/mmmoof1/vae-ft-mse-840000-ema-pruned/blob/main/vae-ft-mse-840000-ema-pruned.safetensors">vae-ft-mse-840000-ema-pruned</a>（我本人比较习惯使用这个VAE），6、对于分辨率的设置，建议横版：768x512（或960x768），竖版手机尺寸：512x768（或768x960），得到以上图片之后，可后续使用高清修复（重绘幅度0.3-0.4，放大算法R-ESRGAN 4x+）放大2倍，以增加丰富图片细节；当然，你也可以选择不用高清修复，使用Tiled Diffusion + Tiled VAE放大手段，那样会得到更加不错的高质量图片，7、这个模型的泛化性还行，可以在不同底模下出图；比如切换到真实风格或动漫风格大模型下生成一些不错的图片，小伙伴们可以尝试。<h3 id="heading-21">三、其他说明：</h3>1、【水墨风lora】标准简短调用词，示例如下：shuimofeng,1girl,blue dress,stand,&lt;lora:shuimofeng:0.9&gt;,masterpiece,best quality,best aesthetics,highest resolution,8K wallpapers,直接加载lora，添加触发词：shuimofeng，再配合一些提示词即可得到一些水墨风格类型的场景和人物图片，2、推荐的一些其他提示词：tree, brunette, water, long hair, landscape, reflection, architecture, outdoors, east asian architecture, petals, pond, lantern, grass, water, rays, flowers, sunshine, bridge, house, village, river, buildings, butterfly, Dragon, Hanfu, fog, boat等等一些跟中国水墨有关的提示词；3、推荐负面提示词：NSFW,(ng_deepnegative_v1_75t:1.5),(easynegative:1.5),(worst quality:2),(low quality:2),(normal quality:2),lowres,low resolution,bad anatomy,bad hands,extra fingers,missing finger,missing fingers,bad feet,extra feet,((monochrome)),((grayscale)),watermark,number,4、关于此模型的使用问题，欢迎大家贴图反馈；若有关于Stable Diffusion的使用和模型训练（炼丹）的任何问题，可一起交流，大家一起玩AI绘画，共同进步；暗号VX：fei_geer。///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////<h3 id="heading-22">I.Overview of the model:</h3><ul><li>Ink-style lora model, trained from nearly 150 ink-style pictures,</li><li>The generated picture contains Chinese ink elements, and the ink style is mainly reflected in the background and strokes.</li><li>I have tested some parameters and loaded the large model first, and more discoveries are waiting for the friends to dig by themselves...</li></ul><h3 id="heading-23">II.The model use instructions:</h3>1, Recommended lora weight: 0.6-0.9, add trigger word: shuimofeng.2, Recommended to load the base large model: <a target="_blank" rel="ugc" href="https://civarchive.com/models/6424?modelVersionId=11745">ChilloutMix</a>或<a target="_blank" rel="ugc" href="https://civarchive.com/models/7371?modelVersionId=43978">ReV Animated_v1.2.2</a>.3, The number of iteration steps: 25-40, CFG scale: 10-15, Clip skip: 2.4, Recommended sampler: DPM++ SDE Karras(You can also try other sampling methods).5, VAE: <a target="_blank" rel="ugc" href="https://huggingface.co/mmmoof1/vae-ft-mse-840000-ema-pruned/blob/main/vae-ft-mse-840000-ema-pruned.safetensors">vae-ft-mse-840000-ema-pruned</a>(I am more used to using this VAE).6, For the resolution settings, the recommended horizontal version: 768x512 (or 960x768), vertical cell phone size: 512x768 (or 768x960), after getting the above picture, you can subsequently use HD repair (redrawing amplitude 0.3-0.4, zoom algorithm R-ESRGAN 4x +) zoom 2 times to increase the rich picture details; of course, you can also choose Instead of HD restoration, you can use Tiled Diffusion + Tiled VAE to enlarge the image, which will give you a better quality image.7, The generalization of this model is not bad, and it can produce pictures under different base models; for example, switch to a real-style or anime-style large model to generate some good pictures, friends can try.<h3 id="heading-24">III. Other notes:</h3>1, [Ink and wash style lora] standard short calling words, examples are as follows:----------------------------------------------------------------------------------------------shuimofeng,1girl,blue dress,stand,&lt;lora:shuimofeng:0.9&gt;,masterpiece,best quality,best aesthetics,highest resolution,8K wallpapers,----------------------------------------------------------------------------------------------Load lora directly, add the trigger word: shuimofeng, and then cooperate with some prompt words to get some ink-wash style scenes and pictures of characters.2, Some other recommended prompt words:----------------------------------------------------------------------------------------------tree, brunette, water, long hair, landscape, reflection, architecture, outdoors, east asian architecture, petals, pond, lantern, grass, water, rays, flowers, sunshine, bridge, house, village, river, buildings, butterfly, Dragon, Hanfu, fog, boat, etc.Some hints related to Chinese ink painting;----------------------------------------------------------------------------------------------3, Recommended negative cue words:-----------------------------------------------------------------------------------------------NSFW,(ng_deepnegative_v1_75t:1.5),(easynegative:1.5),(worst quality:2),(low quality:2),(normal quality:2),lowres,low resolution,bad anatomy,bad hands,extra fingers,missing finger,missing fingers,bad feet,extra feet,((monochrome)),((grayscale)),watermark,number------------------------------------------------------------------------------------------------4, Regarding the use of this model, everyone is welcome to map feedback;If there are any questions about the use of Stable Diffusion and model training (alchemy), you can communicate together, we play AI painting together and progress together.

Files with SHA256: e45d24f3f2944e9ce4cf5af25f7987aa8bd05e86082199d356fe91c750b996e3