一文讲透ai作画原理技术 - AI绘画每日一帖
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">在上一篇<span style="color: black;">文案</span> <a style="color: black;">ai绘画是什么意思?什么是ai绘画?</a> 中,<span style="color: black;">咱们</span>讲到<span style="color: black;">近期</span>火热的 AI 作画技术是是<span style="color: black;">经过</span>文本描述生成绘画,今天<span style="color: black;">咱们</span>就讲一下这<span style="color: black;">暗地里</span>的ai绘画技术和ai画画原理。</p>
<h2 style="color: black; text-align: left; margin-bottom: 10px;">“大象在天上飞”~ 当 AI <span style="color: black;">起始</span>想象</h2>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">早在 1980s,人工智能的先行者们就在尝试<span style="color: black;">处理</span> AI 识别物体的问题,<span style="color: black;">最后</span>在 2015 年 AI 的识别能力超越了人类水平。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic4.zhimg.com/80/v2-f88ab46294290ead6ae26d0497b4bab3_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">能识别<span style="color: black;">照片</span>中的物体后,<span style="color: black;">火速</span> AI 成功地将这些标签组合成一句话,这<span style="color: black;">便是</span>图像字幕技术(image captioning):<span style="color: black;">经过</span>图像生成对应的一句话描述。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic4.zhimg.com/80/v2-657ce25d83f6e1baf5b58f043e998023_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">这个过程能<span style="color: black;">不可</span>反过来?换言之,能<span style="color: black;">不可</span><span style="color: black;">经过</span><span style="color: black;">照片</span>生成描述<span style="color: black;">照片</span>的一句话呢?</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic2.zhimg.com/80/v2-4f0b93abbf666153c8a4892dc026e775_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">相比从<span style="color: black;">照片</span>生成字幕,这是相当大的挑战,<span style="color: black;">科研</span>者<span style="color: black;">期盼</span> AI 能生成人们前所未见的<span style="color: black;">照片</span>。2016 年,这一设想<span style="color: black;">作为</span>了现实,<span style="color: black;">便是</span>这些 32 * 32 像素的<span style="color: black;">照片</span>。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic1.zhimg.com/80/v2-dcd6549aa226eaafd85dda8895bacef8_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">这为<span style="color: black;">咱们</span>展示了<span style="color: black;">有些</span><span style="color: black;">将来</span>的可能性,而<span style="color: black;">此刻</span>,<span style="color: black;">将来</span>已来!</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic1.zhimg.com/80/v2-408ed98883fa3d02dd1dbf4e6c41cc2c_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<h2 style="color: black; text-align: left; margin-bottom: 10px;">AI = 缝合怪?</h2>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">说到</span> AI 作画,<span style="color: black;">非常多</span>批评者会<span style="color: black;">说到</span> “缝合怪” “抄袭”。<span style="color: black;">咱们</span>可能会假设,当<span style="color: black;">咱们</span>输入 “一只骑摩托车的大熊猫”</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic1.zhimg.com/80/v2-82a5e6f641099553b10745a81339b16c_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">AI 会在数据库里检索 “摩托车”、“大熊猫” 的<span style="color: black;">照片</span>,<span style="color: black;">而后</span>把<span style="color: black;">她们</span>拼在<span style="color: black;">一块</span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic3.zhimg.com/80/v2-7d7da95f3a9e7d9e33e142690137fe0a_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">但<span style="color: black;">实质</span>上并非如此,要<span style="color: black;">认识</span> AI 怎么生成<span style="color: black;">照片</span>,<span style="color: black;">必须</span>先理解 latent space——潜在空间。<span style="color: black;">大众</span>都有自己的身份证号码,前 6 位<span style="color: black;">表率</span>地区、中间 8 位<span style="color: black;">表率</span>生日、后 4 位<span style="color: black;">表率</span>个人其他信息。放到空间上如图所示,这个空间<span style="color: black;">便是</span>「人类潜在空间」。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic3.zhimg.com/80/v2-9686b7533a90ae0505f43d4738b4513a_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">这个空间上相近的人,可能<span style="color: black;">便是</span>生日、地区接近的人。人<span style="color: black;">能够</span>对应为这个空间的一个点,这个空间的一个点<span style="color: black;">亦</span>对应一个人。<span style="color: black;">倘若</span>在空间中我的<span style="color: black;">周边</span>找一个点,对应的人可能跟我非常<span style="color: black;">类似</span>,没准<span style="color: black;">便是</span>我失散<span style="color: black;">数年</span>的兄弟 hh</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">AI <span style="color: black;">便是</span><span style="color: black;">经过</span>学习找到了一个「<span style="color: black;">照片</span>潜在空间」,每张<span style="color: black;">照片</span>都<span style="color: black;">能够</span>对应到其中一个点,相近的两个点可能<span style="color: black;">便是</span>内容、风格<span style="color: black;">类似</span>的<span style="color: black;">照片</span>。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic3.zhimg.com/80/v2-558d388bf418ff0ed679f26091acbcf6_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">因此</span>这个空间中有一个区域是 “大熊猫区”,一个区域是 “摩托车区”。提示语 “一只骑摩托车的大熊猫” 会<span style="color: black;">帮忙</span> AI 找到「<span style="color: black;">照片</span>潜在空间」中某个可能<span style="color: black;">位置于</span> “大熊猫区”、“摩托车区” 交汇处的点。AI 再把这个点<span style="color: black;">经过</span>某种方式「生成」一张<span style="color: black;">照片</span>,这种方式<span style="color: black;">便是</span>大名鼎鼎的 “Diffusion”。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://pic1.zhimg.com/80/v2-dc8de629432f0df0f5f97f90f465d8dc_720w.webp" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">至于 AI 是怎么<span style="color: black;">经过</span> prompt(提示语)找到「<span style="color: black;">照片</span>潜在空间」中对应的点,再把这个点生成一张<span style="color: black;">照片</span>,敬请关注<a style="color: black;">ai绘画是怎么画的?ai绘画算法揭秘</a><span style="color: black;">照片</span>引用:</p>“大象在天上飞”~ 当 AI <span style="color: black;">起始</span>想象<a style="color: black;"><span style="color: black;">https://</span><span style="color: black;">github.com/floydhub/ima</span><span style="color: black;">ge-classification-templat</span></a>e<a style="color: black;"><span style="color: black;">https://</span><span style="color: black;">towardsdatascience.com/</span><span style="color: black;">image-captioning-in-deep-learning-9cd23fb4d8d2</span></a><a style="color: black;"><span style="color: black;">https://</span><span style="color: black;">arxiv.org/pdf/1511.0279</span><span style="color: black;">3.pdf</span></a><a style="color: black;"><span style="color: black;">https://www.</span><span style="color: black;">youtube.com/watch?</span><span style="color: black;">v=SVcs</span></a>DDABEkM AI = 缝合怪?<a style="color: black;">画宇宙 - 人工智能 AI 作画网站</a><a style="color: black;"><span style="color: black;">https://</span><span style="color: black;">joeschmoe.io/api/v1/ran</span><span style="color: black;">dom</span></a><a style="color: black;"><span style="color: black;">https://</span><span style="color: black;">medium.com/mlearning-ai</span><span style="color: black;">/latent-space-representation-a-hands-on-tutorial-on</span></a>-autoencoders-in-tensorflow-57735a1c0f3f<a style="color: black;"><span style="color: black;">https://</span><span style="color: black;">unsplash.com/</span></a>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">【原创】</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">作者:倒立的BOB</p>原文请参考:<a style="color: black;">一文讲透ai作画原理技术</a>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">更加多</span>精彩内容请<span style="color: black;">拜访</span> ~</p><a style="color: black;">画宇宙 - 人工智能 AI 作画网站</a>
我完全同意你的观点,说得太对了。 回顾过去一年,是艰难的一年;展望未来,是辉煌的一年。
页:
[1]