AI技术助力B端创作:运营类3D Banner设计新思路
现今,许多B端设计师在日常设计中难免会遇到各种运营 3D banner 设计需求。在设计过程中,他们不仅需要费尽心思构思各种造型,还要不断进行重复渲染,而有时渲染结果也难以令人满意。本文旨在介绍一种基于 Stable Diffusion 混合 AI 的B端 3D Banner 设计方法和流程,可供任何对该领域感兴趣的人进行实验,创作出各类B端模型。
部署 Stable Diffusion 流程
云端部署
拉取镜像(10-15min)
sudo docker pull gpulab.tencentcloudcr.com/ai/stable-diffusion:1.0.8
启动容器,完成部署(1min)
sudo docker run -itd --gpus=all --network=host --device=/dev/dri --group-add=video --ipc=host --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --name=stable-diffusion gpulab.tencentcloudcr.com/ai/stable-diffusion:1.0.8 | xargs sudo docker logs --follow
sudo docker restart stable-diffusion | xargs sudo docker logs --follow
本地安装 Stable Diffusion
| 优势 | 劣势 | |
|---|---|---|
| 云端安装 | 根据情况选择硬件,成本可控;即开即用 | 需要手动安装插件;部署存在一定门槛 |
| 本地安装 | 预装包插件较完整 | 极度依赖本地硬件;成本高 |
3D banner 模型训练流程
收集设计素材,准备训练集
图片处理与裁切
图片预处理操作(手动为图片添加描述)
使用 Dreambooth 进行训练
到 dreambooth 选项卡中,选择刚刚创建的模型:tencent cloud_banner
Instance prompt:输入的 tenentcloud(这个名字不要和现实中存在的常见词语冲突)
Dataset Directory:填写你输出的图片和文本的目录
Class Prompt:填写icon/或者品类
Classification Dataset Directory 和 Total Number of Class/Reg Images 的参数根据自己的需要来填写,例如:40
Learning Rate 和 Training Steps 这两个选项都是决定训练强度的,数字越大,学习效果越强,学习效果越强,就越容易过拟合,但是过低又会欠拟合
Train Wizard 如果是训练人物模型的可以选择 lora,不是的话可以不用选择
点击"Generate Ckpt",大概4个小时之后就可以炼丹成功(根据显卡配置测算时间,2080T大概时间6小时,3080T大概时间4小时)
设计师生产流程
关键词写法:内容,风格,质量,视角四个方向填写关键词
以“服务器”为例。正面关键词:A server, a round object with blue center and top white center, top with light blue center and white center, white background, very high quality 3D ICON. The model is divided into two parts, top and bottom. The bottom is a white metal cube with a slightly glassy texture. There are metal screws at all four corners. The screws are very small. There is only one main object in the scene, the object is on the right side of the screen, and the camera is an isometric perspective. X-axis is -20°, y-axis is 45°, z-axis is 0°, masterpiece, best quality, high resolution ;负向描述:nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,fuzzy structure
采样迭代步数:20-30(不是越高越好,过高也会出现抽象的内容)
生成数量:跟随自己的电脑配置来填写参数,配置好填写数量高,配置低填写低
宽度/高度:512*512
最后的生成效果(我们挑选了一些生成较好的效果)
以“AI 大脑”为例。正面关键词:A brain, a round object with blue center and top white center, top with light blue center and white center, white background, very high quality 3D ICON. The model is divided into two parts, top and bottom. The bottom is a white metal cube with a slightly glassy texture. There are metal screws at all four corners. The screws are very small. There is only one main object in the scene, the object is on the right side of the screen, and the camera is an isometric perspective. X-axis is -20°, y-axis is 45°, z-axis is 0°, masterpiece, best quality, high resolution <lora:DDicon:1>;负向描述:nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,fuzzy structure
采样迭代步数:20-30(不是越高越好,过高也会出现抽象的内容)
生成数量:跟随自己的电脑配置来填写参数,配置好填写数量高,配置低填写低
宽度/高度:512*512
生成结果: