u-net分层融合的，主要是类3d效果，可以是人物，动物，衣服，各种形象，场景，背景，风景等等例图风格种类比较丰富，大家可以看一看，也可以自然语言描述该底膜主要是追求各种方面的泛化性，当然lora的泛化性也不错，同时也感谢各位模型作者们有那么多好的模型，建议高清放大，权重随意，提示词也随意，采样器随意放大算法我常用4x-ultrasharp和8x_nmkd-superscale_150000_g可以使用dynamic thresholding，或者直接cfg从低到高都可以尝试，有不同的小惊喜vae记得用哈，不然画面比较灰，hires steps不建议0哈~~The layered fusion of u-net mainly involves 3D like effects, which can include characters, animals, clothing, various images, scenes, backgrounds, landscapes, and so onThere are various styles in the sample images, which can be taken a look or described in natural languageThe bottom film mainly pursues various aspects of generalization, and of course, Lora's generalization is also good,Thank you also to the model authors for having so many good models,Suggest high-definition amplification, with arbitrary weights, prompt words, and samplersI often use 4x ultra sharp and 8x amplification algorithms_ Nmkd superscale_ 150000_ GYou can use dynamic thresholding or simply try cfg from low to high, with different small surprisesVae, remember to use it, otherwise the screen will be gray. Hires steps do not recommend 0~~

2.99Dmix | Class 3D

V2.0主要是对3D效果的增强，人物更加立体，加强了光影的效果另外手部的崩坏几率也降低不少，所有例图中的手部均为一次出，未经过后期修改V2.0版例图用的提示词与V1.0版一模一样，可以对比两者，酌情使用泛化性依旧比较好，人物，特征，场景均有不错的表现脸部修复请使用adetailer.<a target="_blank" rel="ugc" href="https://github.com/Bing-su/adetailer">https://github.com/Bing-su/adetailer</a>可以开启Dynamic Thresholding<a target="_blank" rel="ugc" href="https://github.com/mcmonkeyprojects/sd-dynamic-thresholding">https://github.com/mcmonkeyprojects/sd-dynamic-thresholding</a>lora分层参考：GitHub - hako-mikan/sd-webui-lora-block-weightV2.0 mainly enhances the 3D effect, making characters more three-dimensional and enhancing the effect of light and shadowIn addition, the probability of hand collapse has also been significantly reduced. The hands in all examples are shown in one shot and have not been modified in the later stageThe prompt words used in the V2.0 example are the same as those used in V1.0. You can compare the two and use them as appropriateThe generalization is still quite good, with good performance in characters, features, and scenesPlease use adetailer for facial repair<a target="_blank" rel="ugc" href="https://github.com/Bing-su/adetailer">https://github.com/Bing-su/adetailer</a>Dynamic Thresholding can be enabled<a target="_blank" rel="ugc" href="https://github.com/mcmonkeyprojects/sd-dynamic-thresholding">https://github.com/mcmonkeyprojects/sd-dynamic-thresholding</a>Lora Layered Reference:GitHub - hako mikan/sd webui la block weight

299dmixClass3D_v20.safetensors

v2.0

v1.0

2.99dmix | Type 3d

u-net layered fusion, mainly for 3d effects, can be characters, animals, clothing, various images, scenes, backgrounds, landscapes, etc.The example styles are quite diverse, everyone can take a look, and also be described in natural languageThe base film mainly pursues generalization in various aspects, of course, loa's generalization is also good,Also, thank you to all the model authors for having so many good models,Recommend high-definition enlargement, weights are arbitrarySampling recommendation DPM++SDE, steps 20, CFG 9 or less, recommended width and height 512,680; 640,512, for scene type images recommend 1024 and aboveThe enlargement algorithms I often use are 4x-ultrasharp and 8x_nmkd-superscale_150000_gDynamic thresholding can be used, or CFG can be tried directly from low to high, there are different surprisesRemember to use vae, otherwise the picture may be gray, not recommended 0 for hires steps~~ v2.0 mainly enhances the 3d effect, making it more three-dimensional, and enhances the effect of light and shadowIn addition, the probability of hand collapse is also greatly reduced, all the hands in the example images are one-time outputs, without any post-editingThe prompt words used in the v2.0 version example images are exactly the same as those in the v1.0 version, you can compare the two and use them as appropriateGeneralization is still quite good For facial repair, please use adetailer.<a href="https://github.com/bing-su/adetailer" target="_blank">https://github.com/bing-su/adetailer</a>You can try enabling dynamic thresholding<a href="https://github.com/mcmonkeyprojects/sd-dynamic-thresholding" target="_blank">https://github.com/mcmonkeyprojects/sd-dynamic-thresholding</a>Lora layered reference:<a href="https://github.com/hako-mikan/sd-webui-lora-block-weight" target="_blank">github - hako-mikan/sd-webui-lora-block-weight</a>

Files with SHA256: 9e8d00faed48bebdc592486fb223be78d48318961e9448ffee63371f2d37ad26