<h1 id="heading-409">TL;DR | 太长不看</h1><pre><code>Trigger word is Eve3D. Use img2img for upscaling. Use ADetailer for different face. Don't use AntiBurn. Recommended weight: XL checkpoints at 0.4 ~ 0.6, 1.5 - realistic/3D checkpoints at 0.6 or 0.7 and 1.5 - anime checkpoints at &lt;=0.5. 
触发词：Eve3D。使用图生图来放大图片获得细节。使用ADetailer 来更换脸型。不要使用AntiBurn。建议LoRA权重为:[XL] 0.4!0.6, [1.5 - 写实/3D] 0.6/0.7, [1.5 - 二次元] &lt;=0.5。
Changelog at the end. 更新日志放在最底部。</code></pre><h1 id="heading-410">Intro | 简介</h1>This LoRA spices up the skin / lighting / image vibe. In simple word it adds a dreamy tone to the final image. This is currently still in the alpha test. It fixes the character's face (in 3D/realistic checkpoints) or draws towards 3D feel (in most anime checkpoints). Use ADetailer or LoRA Block Weight to reduce this issue.此为增强皮肤 / 光照 / 氛围的LoRA，可用于不同风格的大模型。简单来说使用后会使图片产生一种朦胧梦幻的3D渲染质感，若在二次元模型中使用，则主要表现为光感增强。LoRA目前仍为早期测试阶段，在3D和写实模型中对脸型影响较大，在二次元模型中会显著增强立体感。请使用ADetailer 来调整脸部。<h1 id="heading-411">1.5Detail | 训练细节（1.5）</h1>The intention was to train this LoRA to understand these 3D concepts better:<pre><code>lens flare / chromatic aberration / film grain / subsurface scattering / caustics</code></pre>In order to achieve that, I have increased the network_dim to 128, and image resolution to 1024*1536. Thanks to <a target="_blank" rel="ugc" href="https://civarchive.com/user/NeoClassicalRibbon">NeoClassicalRibbon</a> for their <a target="_blank" rel="ugc" href="https://note.com/takumi__ncr/n/n21016c358ea5">insights on training Hipoly 3D Model LoRA</a>.The result was pretty acceptable. In my testing this LoRA indeed catches some of the effects mentioned above. I did a performance test on most of the checkpoints I constantly use, you can download the performance spreadsheet <a target="_blank" rel="ugc" href="https://drive.google.com/drive/folders/1-hQ6ZpqayHaCV1f2Vk14R4qJ07yePjKt?usp=sharing">HERE</a>.此LoRA的目的是为了微调以下3D渲染概念：<pre><code>镜头光晕 / 边缘色散 / 胶片颗粒 / 次表面散射 / 焦散</code></pre>为了令模型学习到这种程度的细节，我将LoRA 网络分辨率上调至128（经过两个版本测试的决定），同时使用1024*1536的训练图像分辨率。训练参考了<a target="_blank" rel="ugc" href="https://civarchive.com/user/NeoClassicalRibbon">NeoClassicalRibbon</a> 所编写的 <a target="_blank" rel="ugc" href="https://note.com/takumi__ncr/n/n21016c358ea5">insight on training Hipoly 3D Model LoRA</a>。LoRA得出的结果目前达到了我的基本预期，在我的测试中它对于部分上述概念拥有了更好的理解能力。你可以<a target="_blank" rel="ugc" href="https://drive.google.com/drive/folders/1-hQ6ZpqayHaCV1f2Vk14R4qJ07yePjKt?usp=sharing">在这里下载</a>查看测试图表。<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3d5ff4b7-cee5-4b5a-a328-4e58e0b74d48/width=525/3d5ff4b7-cee5-4b5a-a328-4e58e0b74d48.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e6d33437-084b-4494-a975-56899f8eb3db/width=525/e6d33437-084b-4494-a975-56899f8eb3db.jpeg" /><h1 id="heading-412">Usage Guide | 使用说明</h1><h3 id="heading-413">1. Activating | 启用</h3>LoRA weight: 0.6-0.8 (below 0.5 for anime checkpoints)Main trigger: Eve3D,Then pick: lens_flare, chromatic_aberration, film_grain, subsurface_scattering, high_contrast, caustics.Not all of them are required, experiment as you see fit. I personally like to keep them all.For this LoRA you don't really need LBW to control weight. Keep in mind that if you wish to generate different faces, you're better off using ADetailer.Images from text2img will most definitely look blurry and "broken". It's fine, a lot of checkpoints have this problem even without any LoRA. Think of it as a way of laying down the basic composition. All the details will appear after upscaling.使用0.6-0.8的LoRA权重，在二次元模型中使用时可以调低至0.5或以下。主要触发词为：Eve3D。然后，你可以选择下列提示词：lens_flare, chromatic_aberration, film_grain, subsurface_scattering, dim_lighting, high_contrast, caustics,不是所有提示词都需要填写，只是我喜欢全部加入。这个LoRA不一定需要LBW控制区块。在文生图时，出来的结果可能看上去像是模糊、损坏的图片，尤其是生成全身图时。对于大部分大模型来说，这是正常现象（因为训练使用的图片分辨率远超1.5文生图所能使用的），只需要通过图生图即可得到正确的结果。<h3 id="heading-414">2. Upscaling | 扩大图片</h3>Use img2img upscale, if you don't, your image will look blurry and bad, and all of the details won't kick in. Below are parameters I use.text2img Steps:32 | Sampler: DPM++ 2M Karras or Eular a | Resolution: 512*768 | CFG: 7 | Hires. fix and Refiner: OFF | Anti Burn: OFFimg2img Steps:48 | Sampler: DPM++ 2M Karras or Eular a | Resolution: 1024*1536 (or higher) | CFG: 7 | Denoising strength: &lt;=0.5 | Hires. fix and Refiner: OFF | Anti Burn: OFF请使用图生图来放大图片和还原细节。以下是我习惯使用的参数，文生图 步数:32 | 采样器: DPM++ 2M Karras or Eular a | 分辨率: 512*768 | CFG: 7 | Hires. fix and Refiner: OFF | Anti Burn: OFF图生图 步数:48 | 采样器: DPM++ 2M Karras or Eular a | 分辨率: 1024*1536 (or higher) | CFG: 7 | 重绘幅度: &lt;=0.5 | Hires. fix and Refiner: OFF | Anti Burn: OFF<h1 id="heading-415">Limitation | 局限</h1><h3 id="heading-344">1.5 LoRA</h3>1. LoRA have strong influence on face. Well at least it's not ugly,2. Film grain doesn't work as intended. Some checkpoints will have hard time understanding these concepts. Some checkpoints misplace "film_grain" as dust on lens. In general the film grain effect is very subtle,3. Text2img quality seems "degraded". Therefore img2img upscale is necessary for this LoRA to work,4. The lens flare effect only occurs under extreme/strong lighting,<s>5. Using "back_lighting" in prompt will sometimes result in character turning back against camera,</s> (problem might be within the prompt itself)6. Have room for captions/images improvements.At current state of this LoRA I would suggest you use this LoRA on anime checkpoints.<h3 id="heading-345">XL LoRA</h3>1. caustic kind of doesn't work. Checkpoints tend to misplace caustic effect as "hologram on clothes",2. Minor overfitting issue, same as 1.5 version.<h1 id="heading-415">Changelog | 更新日志</h1><ul><li>XL_v0.3: First upload of XL1.0 LoRA.</li></ul><ul><li>v0.4: Minor improvement update. Increased repeats, caption improvement.</li><li>v0.3: Initial Civitai upload. Manual caption correction, removed some unnecessary images, increased network_dim to 128, reduced lr and unet_lr.</li><li>v0.1 and v0.2: Never published. Had serious overfitting issue.</li></ul><hr />If you find this LoRA useful, please share what you've made with it. I would love to see how it performs on different environments. Please consider leaving a like.

Files with SHA256: 2cd02456ec467dd6f30346ab87735c881cf1b655762f47ead39484d07efa8330