Type | |
Stats | 1,089 810 |
Reviews | (166) |
Published | Jun 20, 2023 |
Base Model | |
Trigger Words | hayao miyazaki \(style\) |
Hash | AutoV2 5CC6A6A96E |
训练集：
《红猪》《悬崖上的金鱼姬》选取素材
使用金字塔噪声和梯度累加步数
食用参数：
分辨率：见预览图
触发词：hayao miyazaki \(style\)，可加可不加
模型：AnythingV5_v5PrtRE
采样器：DPM++ 2M SDE Karras（其余默认）
权重：0.5~1
这是一个仿宫崎骏画风的LoRA，它的收敛貌情况貌似很好，在经过测试后无拟合问题，测试中只使用了AnythingV5_v5PrtRE，在其它模型上的表现尚未测试，没有对tag做什么人工干涉完全依赖WD1.4打标（懒狗一条），在测试中没有对人物做过多描述，那么1boy的时候大概率是出来宗介样式的。
对于画风LoRA来说，我认为它已经足够了，可能不是能够很好的还原出宫崎骏的风格（我在取材的时候发现动漫中还有油画风，我对宫崎骏不太了解，所以不敢确定还原度如何），（也有人说像鸟山明
Training Set:
Material selected from "Porco Rosso" and "Princess Mononoke."
Utilizing Perlin Noise and Gradient Accumulation Steps.
Parameters:
Resolution: Refer to preview image.
Trigger: hayao miyazaki (style), optional but can be included.
Model: AnythingV5_v5PrtRE.
Sampler: DPM++ 2M SDE Karras (default for the rest).
Weights: 0.5-1.
This is a LoRA (Language to Realistic Artwork) inspired by Hayao Miyazaki's style. It appears to have good convergence based on testing and does not exhibit any overfitting issues. The testing was conducted solely using the AnythingV5_v5PrtRE model, and its performance on other models has not been tested yet. No manual intervention was made in terms of tags; it relied entirely on WD1.4 labeling (I'm a lazy dog). The testing did not involve excessive descriptions of characters, so when generating 1 boy, it is highly likely to resemble the style of Zenko.
Regarding the LoRA's art style, I believe it is already quite satisfactory. It may not perfectly replicate Hayao Miyazaki's style (I discovered some oil painting elements in the anime material, and I'm not very familiar with Miyazaki's work, so I can't be certain about the degree of fidelity). Some people even say it resembles Akira Toriyama's style.
QQ：2085690574