Pokkuti-StyleV1 BY (PONY)

中文：基于PONY模型打造的新pokkuti(style)的LORA，训练尺寸512*768，挑选了145张关于pokkuti图片，训练了17000步。使得LORA整体接近画师的风格！

此处版本为V1，因为上个礼拜才从手术室出院，在家里做的很仓促来不及多修改。不过跑了十几张图片发现效果还不错，大家可以尝试一下！这段时间肠胃不太好，得了急性肠胃炎和消化道疾病，投入AI时间并不多😕

在2025年8月20日入手了全新RTX5080FE显卡后，相比原先3070显卡快了不是一点半点。

从今以后我发布的所有模型将基于FP16打造！

STEP推荐20以上，EULE-A，CFG最佳效果为5.5其次是6再其次是5，放大算法推荐使用(R-ESRGAN 4x+ Anime6B)，没有关键词，LORA添加进去即可。LORA权重为0.5-0.8,，多次尝试0.6是最好的！

后续几天我会出V2版本，将基于768*1152大小打造的迭代版本！还请敬请期待！～(≧∇≦)ﾉ

（推荐使用这个模型进行AI画图，因为我是用这个模型来训练LORA的！mistoonAnime_ponyAlpha.safetensors）

English：This LoRA model is based on the PONY architecture and features the distinctive style of artist Pokkuti. It was trained on 145 selected Pokkuti-style images at a resolution of 512×768 over 17,000 steps, allowing the model to closely capture the original artist’s visual language.

This is marked as V1 — and it was put together in a bit of a rush after I was discharged from the hospital last week. Despite that, I’ve generated a dozen or so test images and the results are already quite promising. Feel free to give it a try! Lately, I’ve been dealing with acute gastroenteritis and some digestive issues, so I haven’t had as much time to devote to AI as I’d like. 😕

That said, after upgrading to the new RTX 5080 FE on August 20, 2025, I’ve seen a huge jump in performance compared to my old RTX 3070.

Moving forward, all models I release will be built using FP16 precision.

For best results with this LoRA:

Use 20+ steps

Sampler: Euler A

CFG Scale: 5.5 (followed by 6, then 5)

Upscaler: R-ESRGAN 4x+ Anime6B recommended

No specific prompts required — just apply the LoRA

LoRA weight: 0.5–0.8 (0.6 works best in most cases)

A V2 update is already in the works, trained at a higher resolution of 768×1152. Stay tuned!～(≧∇≦)ﾉ

（If you're generating images with this LoRA, I definitely recommend using the mistoonAnime_ponyAlpha.safetensors model. It's what I used for training and will best activate the LoRA's style!）