WEBVTT
Kind: captions
Language: zh-Hans

00:00:00.168 --> 00:00:01.736
大家好，我是可波

00:00:01.736 --> 00:00:04.368
ChatGPT 做图功能这次大改版了

00:00:04.368 --> 00:00:05.136
你知道吗？

00:00:05.136 --> 00:00:06.856
每个礼拜光是 ChatGPT 加上

00:00:06.856 --> 00:00:09.220
API 就生出超过三十亿张图

00:00:09.839 --> 00:00:12.307
这次 OpenAI 发表 Images 2.5

00:00:12.307 --> 00:00:14.654
所有方案都有，免费版也能用

00:00:14.654 --> 00:00:16.509
手机、电脑、网页版都有

00:00:17.445 --> 00:00:19.211
这一代生图模型有哪些进步呢？

00:00:19.815 --> 00:00:21.565
OpenAI 在 X 上面就讲了四点——

00:00:22.009 --> 00:00:24.206
速度更快、可以更像原图

00:00:24.206 --> 00:00:25.745
而且改很多次也不会走样

00:00:26.425 --> 00:00:27.957
还能直接在图上面留言

00:00:27.957 --> 00:00:29.802
圈一个地方告诉它你想改什么

00:00:30.635 --> 00:00:33.277
其实每次改版都会讲画质变好

00:00:33.277 --> 00:00:34.156
不意外

00:00:34.156 --> 00:00:36.907
但这次有一件事是以前真的做

00:00:36.907 --> 00:00:38.532
不到的——你叫它改一个地方

00:00:38.532 --> 00:00:40.166
它就真的只改那里

00:00:40.166 --> 00:00:41.166
其他东西不动

00:00:42.210 --> 00:00:44.291
接下来我们会整理四个以前用

00:00:44.291 --> 00:00:46.075
AI 做图最卡的地方

00:00:46.075 --> 00:00:47.700
一个一个看它解到什么程度

00:00:48.169 --> 00:00:49.822
还有这次推出的三个新工具

00:00:50.452 --> 00:00:53.382
画草图、在图上留言、还有模板

00:00:54.209 --> 00:00:56.119
先把官方的公告摊开来看一遍

00:00:56.738 --> 00:00:58.630
这次总共出了两个新模型

00:00:58.630 --> 00:01:00.068
三个新工具，先听 OpenAI

00:01:00.068 --> 00:01:02.146
自己的工程师怎么介绍

00:01:02.798 --> 00:01:05.471
我们今天推出 GPT Image 2.5

00:01:05.471 --> 00:01:06.540
两款新的生图模型

00:01:06.540 --> 00:01:08.678
Sunburst 和 Flare

00:01:08.678 --> 00:01:11.498
Sunburst 是我们最强的生图模型

00:01:11.498 --> 00:01:14.718
画质大幅进步，细节更清楚

00:01:14.718 --> 00:01:17.938
光线更自然，质感也更细腻

00:01:17.938 --> 00:01:19.432
用 Sunburst

00:01:19.432 --> 00:01:21.374
你想像中的画面能做得更精致

00:01:21.374 --> 00:01:22.718
风格和细节都更到位

00:01:22.718 --> 00:01:24.939
你也能分好几次慢慢改

00:01:24.939 --> 00:01:27.605
像是试梦想客厅的各种摆法

00:01:27.605 --> 00:01:30.048
想留的细节更不容易跑掉

00:01:30.048 --> 00:01:32.021
它也更能照详细的指示来画

00:01:32.021 --> 00:01:34.158
连人物用哪只手写字都能照办

00:01:34.323 --> 00:01:36.099
刚刚视频里讲了两个名字

00:01:36.099 --> 00:01:37.878
Sunburst 跟 Flare

00:01:37.878 --> 00:01:39.377
Sunburst 画质最好

00:01:39.377 --> 00:01:40.415
Flare 速度最快

00:01:41.064 --> 00:01:43.176
不过你在 ChatGPT 里面不用管

00:01:43.176 --> 00:01:44.926
这两个名字，它缺省就是 2.5

00:01:44.926 --> 00:01:45.926
直接打字就好

00:01:46.288 --> 00:01:48.016
Sunburst 跟 Flare

00:01:48.016 --> 00:01:49.829
是给写程序的人在 API 那边选的

00:01:50.658 --> 00:01:52.557
改的东西说起来蛮多的

00:01:52.557 --> 00:01:54.536
分两堆来看比较清楚

00:01:54.536 --> 00:01:56.783
模型本身——画面更细

00:01:56.783 --> 00:01:59.513
改完更像、改好几轮不走样

00:01:59.513 --> 00:02:00.513
速度快一半

00:02:00.872 --> 00:02:04.377
然后 ChatGPT 里面多了三个工具

00:02:04.377 --> 00:02:06.002
画草图、可以在图上留言

00:02:06.002 --> 00:02:09.071
还有添加了很多可以直接套用的模板

00:02:10.487 --> 00:02:12.666
有一个第三方的数字可以参考

00:02:12.666 --> 00:02:14.553
Arena 是一个盲测网站

00:02:14.553 --> 00:02:16.495
让一般人投票选哪张图比较好

00:02:17.137 --> 00:02:18.637
文生图排行榜，前三名全是

00:02:18.637 --> 00:02:21.903
OpenAI 的——第一名 Sunburst

00:02:21.903 --> 00:02:23.008
第二名 Flare

00:02:23.008 --> 00:02:24.605
第三名才是自家上一代

00:02:25.419 --> 00:02:27.600
接下来就直接讲以前 AI

00:02:27.600 --> 00:02:29.690
做图最卡的第一件事

00:02:29.690 --> 00:02:31.668
这个问题大家应该都碰过

00:02:31.668 --> 00:02:33.151
你拿一张既有的图

00:02:33.151 --> 00:02:35.511
想请 AI 只改其中一个地方

00:02:35.511 --> 00:02:37.037
这个之前是几乎做不到的

00:02:37.705 --> 00:02:39.114
比如说换个发型

00:02:39.114 --> 00:02:40.800
结果连脸都不一样了

00:02:40.800 --> 00:02:43.027
或者你想把桌上一杯咖啡拿掉

00:02:43.027 --> 00:02:45.435
它连旁边的光线和摆设都重来

00:02:46.254 --> 00:02:47.254
为什么会这样？

00:02:47.607 --> 00:02:50.535
说穿了就是因为以前的做法是

00:02:50.535 --> 00:02:52.572
“参考这张图，重新画一张”

00:02:53.637 --> 00:02:55.537
每改一次就是重新赌一次

00:02:56.239 --> 00:02:58.449
赌这次画出来到底像不像

00:02:59.948 --> 00:03:01.420
比起画质变好

00:03:01.420 --> 00:03:02.547
这件事其实更重要

00:03:03.585 --> 00:03:05.853
因为大部分人丢进去的不是随便

00:03:06.467 --> 00:03:07.989
一张图——是自己的照片

00:03:07.989 --> 00:03:10.028
自己养的狗、要拿去卖的商品

00:03:11.106 --> 00:03:12.356
你要的不是画得漂亮

00:03:12.900 --> 00:03:14.270
是改完还认得出来

00:03:15.088 --> 00:03:16.386
Images 2.5

00:03:16.386 --> 00:03:18.381
说它这次真的把这件事做好了

00:03:18.987 --> 00:03:20.223
官方的说法是

00:03:20.223 --> 00:03:22.786
模型现在更清楚哪些地方不要动

00:03:23.803 --> 00:03:25.053
公告页上有几组对照

00:03:25.543 --> 00:03:27.637
一只狗狗照片换上小礼服

00:03:27.637 --> 00:03:29.774
狗本身完全没变

00:03:29.774 --> 00:03:32.274
一张车票只换了目的地的城市

00:03:32.274 --> 00:03:34.652
版面上其他的字一个都没跑

00:03:35.861 --> 00:03:37.212
光看官方的不够嘛

00:03:37.212 --> 00:03:38.962
我们自己拿可波试了一下——先生

00:03:39.174 --> 00:03:41.995
一张可波坐在沙发上用笔电的图

00:03:41.995 --> 00:03:43.876
然后叫它穿蓝色连帽外套

00:03:43.876 --> 00:03:45.554
再换到海边，再加墨镜

00:03:46.281 --> 00:03:49.381
改了三轮，每一次只动一个地方

00:03:49.381 --> 00:03:51.663
可波的脸从头到尾都没有变

00:03:52.510 --> 00:03:54.360
接下来第二个问题其实

00:03:54.360 --> 00:03:55.854
跟第一个有关

00:03:55.854 --> 00:03:59.457
那就是就算每次改完不会整张

00:03:59.457 --> 00:04:01.207
重画好了，可是经常改到第三轮

00:04:01.207 --> 00:04:02.631
第四轮的时候

00:04:02.631 --> 00:04:05.213
回头看已经跟第一轮确认过的东西

00:04:05.213 --> 00:04:06.449
差了十万八千里

00:04:07.126 --> 00:04:09.015
而且画面会越改越糊

00:04:09.015 --> 00:04:10.574
细节一轮一轮在流失

00:04:11.852 --> 00:04:13.207
Images 2.5

00:04:13.207 --> 00:04:14.981
说它做到了多轮编辑的一致性——

00:04:15.597 --> 00:04:17.446
每一轮接着上一轮改

00:04:17.446 --> 00:04:19.582
前面确认过的东西会留着

00:04:19.582 --> 00:04:21.225
画质也不会越改越差

00:04:22.057 --> 00:04:24.409
公告页上有三组短片可以看

00:04:24.409 --> 00:04:25.684
乍看像动画

00:04:25.684 --> 00:04:28.663
其实是同一张图改很多次串

00:04:28.663 --> 00:04:30.459
起来的——蜡烛一根一根点上去

00:04:30.459 --> 00:04:31.459
蛋糕没有变

00:04:31.598 --> 00:04:33.373
方块一格一格转过去

00:04:33.373 --> 00:04:34.373
光影跟着动

00:04:34.683 --> 00:04:37.283
一张旅游信息图一站一站加上去

00:04:37.283 --> 00:04:39.003
前面加过的每一站都还在

00:04:39.823 --> 00:04:41.390
做到这件事之后

00:04:41.390 --> 00:04:43.763
社群这几天已经有人做出以前做

00:04:43.763 --> 00:04:44.763
不出来的东西

00:04:45.063 --> 00:04:47.859
有人用三十六张图串成定格动画

00:04:47.859 --> 00:04:50.265
有人一句话出整套分镜

00:04:50.265 --> 00:04:51.515
每一张的角色长一样

00:04:52.113 --> 00:04:53.671
还有人做出可以一页

00:04:53.671 --> 00:04:54.671
一页翻的小书

00:04:55.045 --> 00:04:57.096
这些东西有一个共同前提——

00:04:57.096 --> 00:04:59.908
每一轮的结果要长得像上一轮

00:05:01.160 --> 00:05:02.410
我们自己也试了一下

00:05:02.818 --> 00:05:05.581
前面可波那张图改到第四轮加雪景

00:05:06.183 --> 00:05:07.950
第五轮加红毛帽再打上字

00:05:08.648 --> 00:05:10.776
回头看跟第一轮还是同一只狗

00:05:11.740 --> 00:05:14.641
五轮下来六次生成全部成功

00:05:14.641 --> 00:05:15.536
脸没有走样

00:05:16.356 --> 00:05:17.559
前面两个问题

00:05:17.559 --> 00:05:18.942
模型变好就能解决

00:05:19.599 --> 00:05:22.355
第三个问题不一样

00:05:22.858 --> 00:05:24.608
问题出在人这边——你想要的画面

00:05:24.608 --> 00:05:26.781
常常只在你脑中，AI 看不到

00:05:26.781 --> 00:05:27.959
你也讲不清楚

00:05:29.079 --> 00:05:31.039
你打了一整段文本描述

00:05:31.039 --> 00:05:32.039
它还是摆错位置

00:05:32.964 --> 00:05:35.053
或者图上有六盆植物

00:05:35.053 --> 00:05:36.053
你想改其中一盆

00:05:36.634 --> 00:05:38.360
你根本讲不出来到底是哪一盆

00:05:39.583 --> 00:05:41.782
这次 ChatGPT 多了三个工具

00:05:41.782 --> 00:05:42.907
就是要解决这件事

00:05:43.718 --> 00:05:45.888
三个分别是——画的

00:05:45.888 --> 00:05:48.756
用 sketch 画草图；圈的

00:05:48.756 --> 00:05:51.648
可以在图上留言；选的

00:05:51.648 --> 00:05:53.431
有大量模板可以让你先选

00:05:54.246 --> 00:05:56.185
第一个，画草图

00:05:56.185 --> 00:05:57.185
叫做 Sketch

00:05:57.647 --> 00:06:00.011
在对话框里打 @Sketch

00:06:00.011 --> 00:06:01.551
会跳出一块画布

00:06:01.551 --> 00:06:04.062
你在上面随便画，画得丑没关系

00:06:04.062 --> 00:06:05.775
它看的不是你画得好不好

00:06:05.775 --> 00:06:07.193
它看的是东西摆在哪里

00:06:07.996 --> 00:06:09.653
画完再加一句文本描述

00:06:09.653 --> 00:06:11.679
告诉它你要什么风格、什么细节

00:06:12.518 --> 00:06:15.827
一个重点要记住——草图管位置

00:06:15.827 --> 00:06:16.874
文本管长相

00:06:17.651 --> 00:06:19.861
你画一坨圆的东西放在左边

00:06:19.861 --> 00:06:21.180
旁边写“螃蟹”

00:06:21.180 --> 00:06:22.863
它就知道螃蟹放左边

00:06:23.958 --> 00:06:25.347
不加文本描述的话

00:06:25.347 --> 00:06:26.347
它会自己猜

00:06:26.710 --> 00:06:29.036
有人试过画一个客厅的草图

00:06:29.036 --> 00:06:31.066
沙发跟狗都猜对了

00:06:31.066 --> 00:06:32.357
但电视被它猜成音响

00:06:33.125 --> 00:06:35.144
所以文本描述还是要写

00:06:35.144 --> 00:06:37.320
不写的话结果很看运气

00:06:38.620 --> 00:06:40.279
第二个，在图上留言

00:06:41.118 --> 00:06:42.294
图生出来之后

00:06:42.294 --> 00:06:43.419
上面有一排工具列

00:06:44.078 --> 00:06:46.078
你可以直接在图上圈一个地方

00:06:46.078 --> 00:06:47.197
写一句话

00:06:47.197 --> 00:06:49.117
像是“沙发换深绿色”

00:06:49.908 --> 00:06:51.056
它先不会送出去

00:06:51.056 --> 00:06:53.020
你可以继续圈第二处、第三处

00:06:53.622 --> 00:06:54.866
全部圈完再一起送

00:06:55.494 --> 00:06:56.854
一次改好几个地方

00:06:57.682 --> 00:06:58.806
我们试了一下

00:06:58.806 --> 00:07:01.140
在一张可波的雪景图上面

00:07:01.140 --> 00:07:04.937
圈出两个地方——一个是沙发要换

00:07:04.937 --> 00:07:07.025
深绿色、一个是围上红色围巾——

00:07:07.025 --> 00:07:08.025
然后一起送出去

00:07:08.588 --> 00:07:09.917
两个地方都改了

00:07:09.917 --> 00:07:12.085
其他地方也真的完全没动

00:07:13.210 --> 00:07:14.636
不过它不是每次都准

00:07:15.243 --> 00:07:17.084
有人试过在一堆植物里面圈

00:07:17.084 --> 00:07:18.658
一盆要它枯掉

00:07:18.658 --> 00:07:20.463
结果它弄枯旁边那盆

00:07:20.463 --> 00:07:22.088
所以改完还是要自己看一下

00:07:22.624 --> 00:07:23.819
不过整体来说

00:07:23.819 --> 00:07:26.539
比以前用打字描述位置好用太多了

00:07:27.379 --> 00:07:30.096
同一排还有橡皮擦跟去背

00:07:30.096 --> 00:07:32.694
橡皮擦的做法是你先在图上擦

00:07:32.694 --> 00:07:33.692
它做记号

00:07:33.692 --> 00:07:35.001
送出之后才真的处理

00:07:35.677 --> 00:07:36.427
不过老实说

00:07:36.427 --> 00:07:38.475
橡皮擦还是蛮容易失败的，我们

00:07:38.475 --> 00:07:41.354
擦一台笔电试了两三次都没擦掉

00:07:41.354 --> 00:07:42.854
后来换擦马克杯才成功

00:07:43.490 --> 00:07:44.837
去背就不多说了

00:07:44.837 --> 00:07:46.567
主体清楚的时候效果还不错

00:07:47.376 --> 00:07:48.758
第三个，模板

00:07:49.685 --> 00:07:52.370
以前打开做图就是一个空白对话框

00:07:52.370 --> 00:07:53.495
自己想怎么开口

00:07:53.959 --> 00:07:57.364
现在左边多了一面模板墙——海报

00:07:57.364 --> 00:07:59.085
周边商品、logo

00:07:59.085 --> 00:08:00.996
室内设计，分好几类

00:08:02.014 --> 00:08:03.116
选一个进去

00:08:03.116 --> 00:08:04.781
它不会马上开始画

00:08:04.781 --> 00:08:06.165
它会先问你问题

00:08:06.165 --> 00:08:08.605
像是品牌叫什么、做什么的

00:08:08.605 --> 00:08:09.605
想要什么方向

00:08:10.047 --> 00:08:11.865
回答完它才动手

00:08:11.865 --> 00:08:13.822
而且是照你的回答客制的

00:08:13.822 --> 00:08:15.002
每次做出来不一样

00:08:15.846 --> 00:08:18.129
我们拿周边商品那个模板试

00:08:18.129 --> 00:08:20.652
告诉它可波、AI 内容

00:08:20.652 --> 00:08:21.652
想做钥匙圈

00:08:22.141 --> 00:08:23.631
它问了三个问题

00:08:23.631 --> 00:08:26.500
五十三秒就出了一个可波的钥匙圈

00:08:27.883 --> 00:08:29.845
模板也有不听话的时候

00:08:29.845 --> 00:08:32.424
有网友分享尝试室内设计的模板

00:08:32.424 --> 00:08:34.536
丢了照片跟平面图进去

00:08:34.536 --> 00:08:36.600
第一次只给了一张普通的效果图

00:08:36.600 --> 00:08:38.350
要再讲清楚一次它才真的照着

00:08:38.350 --> 00:08:39.519
平面图来做

00:08:40.183 --> 00:08:41.683
所以如果你遇到失败的话

00:08:41.683 --> 00:08:43.006
也可以多尝试几次看看

00:08:43.851 --> 00:08:45.140
接下来说第四点

00:08:45.931 --> 00:08:48.274
以前用 ChatGPT 做一张图

00:08:48.274 --> 00:08:49.622
要等将近一分钟

00:08:49.622 --> 00:08:51.122
等的时候你什么都不能做

00:08:51.678 --> 00:08:53.373
一张就快一分钟的话

00:08:53.373 --> 00:08:54.873
你不会想叫它一次做六张

00:08:55.836 --> 00:08:58.350
不过这次 OpenAI 表示

00:08:58.350 --> 00:08:59.600
生图速度有明显加快

00:08:59.819 --> 00:09:01.069
我们看他们自己怎么说

00:09:01.716 --> 00:09:03.782
第二款模型 Flare

00:09:03.782 --> 00:09:06.036
是我们目前最快的生图模型

00:09:06.036 --> 00:09:09.556
比 GPT Image 2 快 50% 以上

00:09:09.556 --> 00:09:10.356
质量一样好

00:09:10.356 --> 00:09:12.836
更快试更多点子

00:09:12.836 --> 00:09:14.930
两款模型都支持透明背景

00:09:14.930 --> 00:09:16.452
从信息图表到游戏

00:09:16.452 --> 00:09:18.736
各种精细的素材都做得出来

00:09:18.736 --> 00:09:22.347
要做高端的生图就用 Sunburst

00:09:22.347 --> 00:09:25.156
要快速出好图就用 Flare

00:09:25.156 --> 00:09:27.513
ChatGPT 和 Codex 都能用

00:09:27.513 --> 00:09:28.381
也开放 API

00:09:28.549 --> 00:09:30.554
刚刚视频里有一组数字

00:09:30.554 --> 00:09:31.652
官方示范里

00:09:31.652 --> 00:09:34.932
同一张图用上一代要四十七点八秒

00:09:34.932 --> 00:09:37.638
换成新的 Flare

00:09:37.638 --> 00:09:39.388
只要十四点三秒，快了超过一半

00:09:40.869 --> 00:09:43.591
快了之后，做法就完全不一样了

00:09:43.591 --> 00:09:45.863
有人试过一次给它七组描述

00:09:45.863 --> 00:09:47.113
让它一张一张生出来

00:09:47.611 --> 00:09:50.080
有人写一篇文章需要六张演示文稿配图

00:09:50.080 --> 00:09:51.552
五分钟全部出来了

00:09:52.700 --> 00:09:55.060
一次做很多张不只是省时间

00:09:55.060 --> 00:09:56.927
你可以让它先出三个设计方向

00:09:56.927 --> 00:09:58.740
挑一个喜欢的再做下去

00:09:58.740 --> 00:10:00.314
一张商品照可以一口气变成

00:10:00.314 --> 00:10:01.314
整套商品图

00:10:01.849 --> 00:10:03.452
或者先出分镜确认构图

00:10:03.452 --> 00:10:05.384
确认完再丢给视频生成的工具

00:10:06.212 --> 00:10:07.249
透明背景也有了

00:10:07.896 --> 00:10:10.752
我们拿可波做了一张五格动作图

00:10:10.752 --> 00:10:12.086
要求背景透明

00:10:12.086 --> 00:10:13.586
第一次生出来就是透明的

00:10:14.130 --> 00:10:16.569
拿下来可以直接贴进其他设计里

00:10:17.920 --> 00:10:19.670
如果你会写程序，API 那边两个

00:10:19.670 --> 00:10:23.276
模型怎么选很简单——Flare 快

00:10:23.276 --> 00:10:24.562
适合日常跟大量出图

00:10:25.382 --> 00:10:27.645
Sunburst 慢一点但控制更细

00:10:27.645 --> 00:10:28.645
适合正式的素材

00:10:29.636 --> 00:10:31.848
两个生出来的图都会自动标记

00:10:31.848 --> 00:10:33.256
让人查得到是 AI 做的

00:10:34.096 --> 00:10:36.578
四个以前最烦的问题跑下来

00:10:36.578 --> 00:10:37.684
只改一个地方

00:10:37.684 --> 00:10:40.516
不会整张重画了。改很多轮

00:10:40.516 --> 00:10:41.961
前面确认过的会留着

00:10:42.806 --> 00:10:44.341
脑中的画面讲不清楚

00:10:44.341 --> 00:10:47.118
现在有三种方式可以比划——画的

00:10:47.118 --> 00:10:48.295
圈的、选的

00:10:49.089 --> 00:10:50.548
嫌太慢，快了一半

00:10:51.171 --> 00:10:52.343
可以一整批交出去

00:10:53.200 --> 00:10:55.005
画质变好，每一代都会讲

00:10:55.701 --> 00:10:57.151
这次真正不一样的

00:10:57.151 --> 00:10:59.552
是它终于分得清楚哪里要改

00:10:59.552 --> 00:11:00.552
哪里不要动

00:11:00.920 --> 00:11:03.053
以前用 AI 做图比较像抽奖

00:11:03.053 --> 00:11:04.895
每改一次就重抽一次

00:11:04.895 --> 00:11:06.303
抽到不像就只能重来

00:11:07.033 --> 00:11:09.640
现在终于比较像跟一个听得懂

00:11:09.640 --> 00:11:11.265
话的设计师合作——你圈哪里

00:11:11.265 --> 00:11:12.640
要改什么，它就改什么

00:11:13.356 --> 00:11:15.193
所以如果你之前试过 AI 做图

00:11:15.193 --> 00:11:16.733
觉得改来改去很烦

00:11:16.733 --> 00:11:19.684
这一版值得再试一次。我是可波

00:11:19.684 --> 00:11:21.309
如果觉得这样的内容有帮助

00:11:21.309 --> 00:11:23.422
请你帮我按赞留言分享

00:11:23.422 --> 00:11:24.955
让我制作更多内容

00:11:24.955 --> 00:11:25.705
我们下次见！