步数减至8步!太炸裂了!FLUX提速新纪元!字节开源FLUX Dev加速Hyper Lora,Flux快速出图,GGUF可用!

前言

FLUX Dev Hyper Lora介绍

8 月 26 日字节推出了FLUX Dev的加速版本分别为8步和16步的Hyper Lora,该Hyper Lora的推出将大幅减少图片生成的步骤,可在8步或16步情况下可达到原来的20-30步的效果,将显著提升广大爱好者的绘图效率。

**关于Hyper Lora模型强度的设置,官方给出的建议约为0.125(设置为0.13即可,ComfyUI会四舍五入),引导值设置为3.5。**较低的Lora比例可能意味着在模型微调时,对原始预训练模型的影响较小,有助于保持模型的稳定性,同时又能根据特定任务进行适度的调整。

模型:

所有的AI设计工具,模型和插件,都已经整理好了,👇获取~在这里插入图片描述

8步在线体验地址:

FLUX 效果展示

通过对8步和16的简单对比测试,在4070上8步的时间大概是10s左右,16步大概在17s左右。

Hyper-FLUX.1-dev-8steps-lora

Hyper-FLUX.1-dev-16steps-lora

Hyper-FLUX.1-dev-8steps-lora出图效果

Hyper-FLUX.1-dev-16steps-lora出图效果

FLUX.1-dev-fp8 无Lora16步直出效果

FLUX.1-dev-fp8 无Lora 30步直出效果

虽然8步模型和16步在效果上可能存在一定的折损,但这种折损在实际应用中老徐觉得应该也是可以接受的,16步的效果老徐觉得还是很不错的。

基础工作流地址:

https://www.liblib.art/modelinfo/c58020318bce4e1aa7d95e41dbac9bda

FLUX Dev Hyper Lora效果展示

下面我们来看一看通过使用Hyper-FLUX.1-dev-16steps-lora其他场景的出图效果

圣冬女孩:

(absurd quality,16k,highres,masterpiece:1.2),ultra details,UHD,(photorealistic Realism 16K Quality),(ultra absurd quality, extremely detailed detail, hyper resolution, clear sharp focus, not blurry, Realistic brown_eyes:1.35),(perfect dark_eyeshadows),Depth of field,cinematic light,Lens flare,dslr,best high quality soft lighting,sharp focus captured by Fujifilm XT3,f 5.6,in a dramatic lighting,((view the viewer:1.3)),Draw a masterpiece of art sexy goddess of holy winter,under the sign of the frosty phoenix. he is dressed in ancient biomechanics,an otherworldly exoskeleton,powered by the goddess's inner negative energy,an energy visible to the viewer. Draw moving icy,blue,black,natural ((living)) flames on the futuristic exoskeleton. This sexy outfit has (((many))) bold,sexy leaks that cause the water in the area to freeze rapidly. Draw ((((ice crystals)))) floating in the air and flying (((water bubbles ))). The cold blue and black((flames)) from the armor very quickly jumps to nearby (ice crystals) and jumps ((further)) and (((further))) trying to freeze the (water) droplets in the distance. Draw Bobino's companion flying the icy phoenix bird in a flood of icy terror.,

海报封面:

A movie poster with Inception as the theme. The overall style of the poster is a surreal double exposure effect, incorporating both reality and dream constructs. The background of the poster features a turbulent cityscape overlaid with spiraling fractal patterns representing the layered nature of dreams. The poster's top center reads "Inception", which is the main title of the poster, rendered in angular futuristic font that melts subtly into the background textures. The poster's bottom right corner reades "What Happens In The Dream Stays In The Dream.," which is the subtitle of the movie, using aged parchment text to hint at subconscious revelation, with subtle watermarks suggesting a shifting dream landscape behind the text itself. Text elements on both sides of the 'what happens......' tagline read "Dreamscape architects: Dom Cobb |", and "Their target: Plant a powerful idea in a corporation" displayed in overlapping font-styles; above these elements are smaller tags stating, "*Opening Day: July 16*", "'*A Christopher Nolan Film*", ,and '*Limited IMAX Experience*. ' located around the top half of the poster. The left side of the poster shows a rotating cityscape that is partially dissolved into abstract clockwork mechanics to symbolize mental manipulation, depicted in bright neon colors; on the right, we see an ethereal figure seemingly melting into thin air—Leonardo DiCaprio as Dom Cobb—holding a dream-manipulating device resembling an ornate brass contraption. The figure blends with swirling fractal patterns and a cascading staircase leading into deeper darkness.

女性杂志封面:

A fashion magazine cover with the text "joe" in white letters, a beautiful Korean woman wearing a mint green suit jacket and black top, holding a flower in her hand, poses for a photoshoot. She has brown, long, straight hair in a medium-length style

红色连衣裙女士

A mature woman,tall figure,red dress (dress details),clean quality,super realistic,flawless,sensational legs,high quality,movie grade,master artwork,perfect face,beautiful face,fantasy,fantasy,description,quality,unreal,science fiction,mature temperament,jewelry,diamonds,emeralds,lace,lace,stockings,clothing,black hair,high detail,high quality,fleshy body,urban background,

牡丹花与蝴蝶

Low light photography,a masterpiece of masters,light particles,sparkle view,(riverside:0.8),night,a butterfly on peony,Pollen dancing,Focus on the stamen,Bokeh,best quality,masterpiece,extreme details,no humans,water drop,watermark,National Geographic Magazine Style,High definition,ultra-high resolution,macro photography,Contour,fluffy light,fire,Flames cover wings GMajic,particle,


这里分享给大家一份Adobe大神整理的《AIGC全家桶学习笔记》,相信大家会对AIGC有着更深入、更系统的理解。

有需要的朋友,可以点击下方免费领取!

在这里插入图片描述

AIGC所有方向的学习路线思维导图

这里为大家提供了总的路线图。它的用处就在于,你可以按照上面的知识点去找对应的学习资源,保证自己学得较为全面。如果下面这个学习路线能帮助大家将AI利用到自身工作上去,那么我的使命也就完成了:
在这里插入图片描述

AIGC工具库

AIGC工具库是一个利用人工智能技术来生成应用程序的代码和内容的工具集合,通过使用AIGC工具库,能更加快速,准确的辅助我们学习AIGC
在这里插入图片描述

有需要的朋友,可以点击下方卡片免费领取!

在这里插入图片描述

精品AIGC学习书籍手册

书籍阅读永不过时,阅读AIGC经典书籍可以帮助读者提高技术水平,开拓视野,掌握核心技术,提高解决问题的能力,同时也可以借鉴他人的经验,结合自身案例融会贯通。

在这里插入图片描述

AI绘画视频合集

我们在学习的时候,往往书籍源码难以理解,阅读困难,这时候视频教程教程是就很适合了,生动形象加上案例实战,科学有趣才能更方便的学习下去。

在这里插入图片描述

### 使用 FluxLoRA 进行机器学习模型训练 #### 准备环境与工具 为了能够顺利地使用 FluxLoRA 来进行模型训练,首先需要准备好相应的开发环境。这通常意味着要安装 Python 及其必要的库文件,比如 PyTorch 或 TensorFlow 等深度学习框架。对于特定于 Flux 的情况,则需按照官方文档指导完成 Julia 编程语言及其依赖项的设置。 #### 加载预训练模型并应用LoRA微调 当准备就绪之后,可以从 Hugging Face Hub 下载预先训练好的基础模型作为起点[^2]。接着利用 LoRA 技术对该模型实施低秩适配(low-rank adaptation),即只更新部分参数而非整个网络结构中的所有权重值。这种方法不仅提高了效率而且减少了过拟合的风险。 ```python from peft import LoraConfig, get_peft_model import torch.nn as nn model = ... # Load your base model here. config = LoraConfig( r=8, lora_alpha=32, target_modules=["q", "v"], lora_dropout=0.05, ) peft_model = get_peft_model(model, config) ``` #### 构建数据集用于训练过程 构建适合当前任务的数据集至关重要。如果目标是创建像“黑神话悟空”这样的角色像生成器,则应收集大量与此主题相关的高质量片样本,并将其整理成可用于训练的形式。这些数据应当被划分为训练集、验证集以及测试集三大部分以便后续评估模型性能[^3]。 #### 开始训练流程 一旦上述准备工作全部完成后就可以启动实际的训练环节了。此阶段涉及到定义损失函数(loss function)、优化算法(optimizer algorithm)以及其他超参的选择。值得注意的是,在每次迭代过程中都要保存好最佳版本的模型副本至指定路径下以供将来部署或进一改进之用[^1]。 ```python output_dir = "./ai-toolkit/output" for epoch in range(num_epochs): ... if best_loss > current_loss: best_loss = current_loss checkpoint_path = f"{output_dir}/best_model.pth" torch.save(peft_model.state_dict(), checkpoint_path) ```
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值