Nai3-Flux

The 11.08GB file is the base model, and the 4.04GB zip file contains the VAE and text encoder. If you already have the VAE and text encoder, you only need to download the 11.08GB base model.
11.08G的文件是模型本体，4.04G的文件是vae以及text encoder，如果你已经有vae和text encoder，那仅下载11.08G的模型本体就可以了

This model builds upon Flux.1 Dev and aims to reproduce a specific style from Nai3 on the Flux (Of course, nai3's art style varies depending on the artist tags you choose, and this model only recreates one particular style). I generated around 400 high-quality images using Nai3 as the training set, then trained a Lora on Flux.1 Dev. Finally, I merged this Lora into Flux.1 Dev-Fp8 at a specific ratio, resulting in this model.
这个模型基于Flux.1 Dev，旨在Flux上最大程度的还原nai3的风格(当然，nai3的画风会根据你选择的画师tag而不同，这里只是复刻了nai3的某一种画风)。我使用Nai3生成了400张左右的高质量图片作为训练集，然后在Flux.1 Dev训练了一个Lora，然后将这个lora以一定比例融合进了Flux.1 Dev-Fp8中，最终得到了这个模型。

It’s better suited for creating aesthetically pleasing portraits with simple backgrounds, rather than for producing grand and elaborate scenes.
相比制作宏大、绚丽的场景画，它更适合制作背景较为简单、风格偏唯美的肖像画

***Please make sure to add "nai3_style_for_flux" to stabilize the art style
***请务必添加"nai3_style_for_flux"以稳定画面风格

I recommend adding the following quality prompts. Based on my testing, they can make the art style more closely resemble the style in the training set:
我推荐加上如下这段质量提示词。经过我的测试，它可以让画风更加接近训练集中的画风：
detailed, dramatic lighting, realistic,detailed,natural lighting,photorealistic,authentic,crisp details,breathtaking scenery,(masterpiece:2),best quality,ultra highres,original,extremely detailed,perfect lighting,Cinematic Lighting,highres,photo_(medium),vibrant colors,surreal,trippy,

I recommend setting flux guidence/Distilled CFG Scale around 3.5 and using Euler+Simple
我推荐将flux guidence/Distilled CFG Scale设置在3.5左右，使用Euler simple采样器

***The base model doesn't come with a VAE or text encoder. If you don't have them, you can download the zip file I've provided (it's 4.04GB in size), which includes both. Once downloaded, unzip the file and put "ae.safetensors" in the "VAE" folder, and "t5xxl_fp8_e4m3fn.safetensors" and "clip_l.safetensors" in the "text_encoder" folder('clip' folder in comfy ui).
***base model中不包含vae以及text encoder，如果你没有它们，可以下载我提供的zip压缩包(大小为4.04G的文件)，里面有vae以及text encoder。下载完成后解压缩，然后把ae.safetensors放到VAE文件夹中，把t5xxl_fp8_e4m3fn.safetensors和clip_l.safetensors放到text_encoder文件夹中 (comfy ui就放进clip文件夹中)。