AI 画出来的图通常应用于以下场景：
广告设计:可快速生成各种创意概念图，为广告策划提供灵感和初稿。游戏开发:用于创建游戏场景、角色形象等，提高开发效率。影视制作:辅助生成特效场景、概念设计等。建筑设计:帮助设计师构想建筑外观和内部布局。
By contrast,if you’re teaching a health class where a student develops an app,you might not care if they use generative AI to help write the code.Instead,your focus is on helping students design a health campaign based on healthy habits.You might not have time to teach students to code by hand.You might not care about coding by hand.The app is merely a way for students to demonstrate their understanding of a health standard.If you’re teaching an art class,you might not want AI-generated images but you might embrace AI-generated images in a history class where students work on making infographics to demonstrate their understanding of macroeconomics principles.It might feel like cheating for a student in a film class to use AI for video editing but the AI-generated jump cuts might save loads of time in a science class where students demonstrate their learning in a video.In a film class,it’s critical for students to learn how to edit by hand in order to tell a story.In science,AI-generated jump cuts allow students to create videos quickly so they can focus on the science content.I also want to recognize that some of what students learn can and will become obsolete.I’m pretty sure I didn’t actually need to memorize the state capitols,for example.Which leads me to the next question...But What If We Don’t Need That Skill Anymore?I grew up in an era where teachers were moving away from memorization.We still had to memorize math facts and,for some reason,state capitols.I’ve never visited a state and thought,“Man,I really need to see the capitol.” If I’m Nevada,I’m not like,“Screw Vegas,I’m going to Carson City!”
游戏里的科技树图标实在太多了，根本画不过来……在没有AI画画之前，我都不敢想啥时候能把这些玩意填完。AI画画出来之后，我抱着满腔热情去试，结果发现三个严重的问题：•全TM在画二次元，没有适合的风格，想用到游戏里必须自己炼丹。•画出来的画好多都是美少女看镜头，没有叙事性，没法当icon用。•已有素材几乎全是中世纪大胡子男人呆呆站着，图生图不可行，训练出来的泛化性也很差。一开始的规划是：•画出卡通简笔画风格。•资产条件：有98张人像和8张UI，且人像全是男人。•需要能产出带有该画风的具有一定叙事内容的图像，内容形式一定要多元。我尝试了最开始的Embedding：后来换成CKPT（画和训练集里接近的小人已经不错了，但泛化性还是不理想）：然后是Lora（好！很接近了！）：现在这个版本画人画物画事都很完美了，甚至能从全是大胡子男人的训练集里学会画女人：从中世纪里摘出摩托车：为了引导AI画出前景后景区分明显的画，还专门画了一组引导图。分别是只保留前景、只保留背景、全图共三张图，在Caption里打组：由于训练集中人物朝向太固定，此处额外做了镜像处理。