1
00:00:00,000 --> 00:00:04,000
大家好我是Kate。OpenEye今天推出了GPT Images 2.5。

2
00:00:04,000 --> 00:00:12,000
官方介绍,与Images 2.0相比呢,2.5呢,生成延迟降低非常多。

3
00:00:12,000 --> 00:00:18,000
而且现在是有更好的照明、文理和可识别的主体。

4
00:00:18,000 --> 00:00:21,000
在重复编辑中,也更好地保留细节。

5
00:00:21,000 --> 00:00:25,000
并且我们可以直接在图像上评论,来针对性地更改。

6
00:00:25,000 --> 00:00:27,840
我们可以在GPT里通过Add a Sketch

7
00:00:27,840 --> 00:00:31,460
就可以画一个炒图

8
00:00:31,460 --> 00:00:33,560
然后让GPT生成图片

9
00:00:33,560 --> 00:00:35,340
而且GPT Image页面呢

10
00:00:35,340 --> 00:00:37,700
目前也上新了非常多的模板

11
00:00:37,700 --> 00:00:40,120
OPI对外推出的API呢

12
00:00:40,120 --> 00:00:41,340
有两个模型

13
00:00:41,340 --> 00:00:42,420
一个是Flare

14
00:00:42,420 --> 00:00:44,180
一个是Soundburst

15
00:00:44,180 --> 00:00:46,140
在Arend的榜单上

16
00:00:46,140 --> 00:00:48,980
GPT Image 2.5是霸榜的

17
00:00:48,980 --> 00:00:51,000
无论是文本生成图片

18
00:00:51,000 --> 00:00:52,440
还是图片编辑

19
00:00:52,440 --> 00:00:54,540
或者是多图片编辑

20
00:00:54,540 --> 00:00:56,840
现在来看一下官方的一些案例

21
00:00:56,840 --> 00:00:59,140
左边是未整理的床

22
00:00:59,140 --> 00:01:00,540
2.5编辑后

23
00:01:00,540 --> 00:01:02,680
整个床就变得整齐了

24
00:01:02,680 --> 00:01:07,380
这次官方特别强调2.5它的精确编辑能力

25
00:01:07,380 --> 00:01:11,760
左边可以看到人物经常和背景是有些变动

26
00:01:11,760 --> 00:01:13,960
而优测的效果好了非常多

27
00:01:13,960 --> 00:01:18,260
也可以让一张旅行票据循环更换城市名称

28
00:01:18,260 --> 00:01:21,260
而票据设计和背景保持一致

29
00:01:21,260 --> 00:01:25,180
所以Images 2.5特别适合来做定格动画

30
00:01:25,180 --> 00:01:28,400
官方展示的生日蜡烛的这样一个演示

31
00:01:28,400 --> 00:01:29,620
也非常有意思

32
00:01:29,620 --> 00:01:31,580
这就是定格动画

33
00:01:31,580 --> 00:01:34,940
Images 2.5更擅长理解复杂的视觉指令

34
00:01:34,940 --> 00:01:37,580
我们可以看一下官方展示的一些

35
00:01:37,580 --> 00:01:39,600
不同风格的案例

36
00:01:39,600 --> 00:01:43,340
包括让它来做PPT

37
00:01:43,340 --> 00:01:45,160
也是非常不错的

38
00:01:45,160 --> 00:01:48,400
现在我们看一下X上面的用户精彩案例

39
00:01:48,400 --> 00:01:50,640
比如说这位博主

40
00:01:50,640 --> 00:01:54,820
他就让GBT Image 2.5做定格动画

41
00:01:54,820 --> 00:01:55,740
周一博主呢

42
00:01:55,740 --> 00:01:59,320
他分享了2.5和2.0生成图片的区别

43
00:01:59,320 --> 00:02:02,140
明显可以感知2.5生成的画面呢

44
00:02:02,140 --> 00:02:03,080
质感更好

45
00:02:03,080 --> 00:02:06,540
他也让Image 2.5生成36张图片

46
00:02:06,540 --> 00:02:09,580
然后让它们做成定格动画

47
00:02:09,580 --> 00:02:12,260
还有博主反馈Image 2.5很猛

48
00:02:12,260 --> 00:02:15,260
他用指定大量物体的提示词去生成

49
00:02:15,260 --> 00:02:17,160
发现没有崩块的

50
00:02:17,160 --> 00:02:18,560
全部画进去了

51
00:02:18,560 --> 00:02:19,560
周一博主呢

52
00:02:19,560 --> 00:02:24,200
他说一条提示词就让images2.5做了整套分镜

53
00:02:24,200 --> 00:02:26,900
而这位博主他说除了定格动画

54
00:02:26,900 --> 00:02:30,420
还可以让2.5呢做翻页书

55
00:02:30,420 --> 00:02:33,000
而这位博主他分享images2.5

56
00:02:33,000 --> 00:02:37,020
生成超写实的iPhone创意肖像

57
00:02:37,020 --> 00:02:40,640
而这位博主呢他做了测试三种不同的鱼

58
00:02:40,640 --> 00:02:42,900
他让为GPT生成的这三种鱼呢

59
00:02:42,900 --> 00:02:45,000
还是没有通过他的测试

60
00:02:45,000 --> 00:02:47,480
还有博主让GPT Images呢

61
00:02:47,480 --> 00:02:50,540
把这些文字平铺在一个图上

62
00:02:50,540 --> 00:02:52,100
现在来看我的实测

63
00:02:52,100 --> 00:02:54,680
我上传这样一张图片给GPT

64
00:02:54,680 --> 00:02:58,340
我选择的是它的image下面的一个模板

65
00:02:58,340 --> 00:03:00,660
是做图标的

66
00:03:00,660 --> 00:03:03,700
它呢会先问我一些问题

67
00:03:03,700 --> 00:03:05,880
然后我回答之后

68
00:03:05,880 --> 00:03:07,860
它就生成这样的图标了

69
00:03:07,860 --> 00:03:09,720
怀孕性还是非常好的

70
00:03:09,720 --> 00:03:11,140
而这一页呢

71
00:03:11,140 --> 00:03:13,260
是我通过它的sketch功能

72
00:03:13,260 --> 00:03:15,040
先画了这样的图片

73
00:03:15,040 --> 00:03:18,500
然后提示它生成一个种子拉拉队

74
00:03:18,500 --> 00:03:19,460
更多的测试

75
00:03:19,460 --> 00:03:22,880
我是在GPT里选择Codex生成的

76
00:03:22,880 --> 00:03:24,000
我先问他

77
00:03:24,000 --> 00:03:26,920
你的图片用的是image2.5吗

78
00:03:26,920 --> 00:03:31,140
他说他无法确认底层是否用GPT image2.5

79
00:03:31,140 --> 00:03:35,580
后来我就问他这个工具对应文档有哪些

80
00:03:35,580 --> 00:03:37,620
接着我就将7个提示词

81
00:03:37,620 --> 00:03:39,520
同时发给GPT

82
00:03:39,520 --> 00:03:41,040
让他拼量生成图片

83
00:03:41,040 --> 00:03:42,460
因为他有agent的能力

84
00:03:42,460 --> 00:03:44,920
所以哪怕30个提示词

85
00:03:44,920 --> 00:03:45,940
50个提示词

86
00:03:45,940 --> 00:03:48,000
都可以直接发给Codex

87
00:03:48,000 --> 00:03:49,060
让它去处理

88
00:03:49,060 --> 00:03:51,220
它会一张张的去生成

89
00:03:51,220 --> 00:03:54,160
2.5生成的这样的一个绘本页面

90
00:03:54,160 --> 00:03:55,680
还是非常不错的

91
00:03:55,680 --> 00:03:56,320
接着呢

92
00:03:56,320 --> 00:03:57,540
我将Astra的博文

93
00:03:57,540 --> 00:03:59,020
发给GPT

94
00:03:59,020 --> 00:04:02,080
让它生成6张PPT样式的图片

95
00:04:02,080 --> 00:04:05,100
最终它生成的效果

96
00:04:05,100 --> 00:04:07,820
我个人还是非常认可的

97
00:04:07,820 --> 00:04:09,160
质感非常高级

98
00:04:09,160 --> 00:04:10,300
我在这里呢

99
00:04:10,300 --> 00:04:11,660
选择的都是Astra

100
00:04:11,660 --> 00:04:15,220
它花了五分钟就生成了

101
00:04:15,220 --> 00:04:16,980
如果是GPTImage2

102
00:04:16,980 --> 00:04:19,040
那它生成这六张图

103
00:04:19,040 --> 00:04:20,260
正常情况下

104
00:04:20,260 --> 00:04:23,060
它是要远超过五分钟的

105
00:04:23,060 --> 00:04:24,720
接着我上传这张图片

106
00:04:24,720 --> 00:04:28,580
让GPT来生成一个精灵GIF

107
00:04:28,580 --> 00:04:31,620
它生成的是这样的效果

108
00:04:31,620 --> 00:04:34,040
后来我又提示它背景要透明

109
00:04:34,040 --> 00:04:35,860
并且给兔子加把剑

110
00:04:35,860 --> 00:04:37,900
展示它五件

111
00:04:37,900 --> 00:04:40,380
这是最终生成的效果也不错

112
00:04:40,380 --> 00:04:43,140
而且生成速度也非常快

113
00:04:43,140 --> 00:04:43,720
三分钟

114
00:04:43,720 --> 00:04:46,180
之后我又让他给这个GIF

115
00:04:46,180 --> 00:04:47,880
再加一个趣味文字

116
00:04:47,880 --> 00:04:52,080
最终他生成的还是非常搞笑的

117
00:04:52,080 --> 00:04:54,000
我可以发给我的朋友

118
00:04:54,000 --> 00:04:56,220
Zalangipity基于对我的了解

119
00:04:56,220 --> 00:04:57,960
制作一张少年漫画

120
00:04:57,960 --> 00:05:01,120
并且提示我的名字

121
00:05:01,120 --> 00:05:02,280
CashyCatch

122
00:05:02,280 --> 00:05:03,820
他生成的角色设定

123
00:05:03,820 --> 00:05:05,820
感觉起来还挺酷的

124
00:05:05,820 --> 00:05:08,860
Zalangipity生成六张分镜图片

125
00:05:08,860 --> 00:05:13,020
并生成一套15秒的动态摄影方案

126
00:05:13,020 --> 00:05:15,400
5分钟它就完成了

127
00:05:15,400 --> 00:05:19,220
我看它生成了这些图片质感

128
00:05:19,220 --> 00:05:20,660
图片间的风格

129
00:05:20,660 --> 00:05:22,680
还是非常像的

130
00:05:22,680 --> 00:05:24,220
刚刚我有介绍

131
00:05:24,220 --> 00:05:26,260
AXA上有用户分享

132
00:05:26,260 --> 00:05:28,380
让Images 2.5生成翻页书

133
00:05:28,380 --> 00:05:30,540
所以我这里也参考了它的思路

134
00:05:30,540 --> 00:05:32,940
这是它最终生成的翻页书

135
00:05:32,940 --> 00:05:36,260
那我可以点击右侧的沉浸阅读

136
00:05:36,260 --> 00:05:39,180
通过左右按键

137
00:05:39,180 --> 00:05:42,260
来看它生成的效果

138
00:05:42,260 --> 00:05:43,340
在Codex里呢

139
00:05:43,340 --> 00:05:44,900
我们就可以很方便的

140
00:05:44,900 --> 00:05:47,020
将我们的想法提出

141
00:05:47,020 --> 00:05:48,140
由一张图片

142
00:05:48,140 --> 00:05:51,220
最终能做出这样的外面页面

143
00:05:51,220 --> 00:05:54,620
它的一致性保持的还是非常好的

144
00:05:54,620 --> 00:05:56,000
前两天我做了

145
00:05:56,000 --> 00:05:58,560
悟空大闹翻桃园的Blender

146
00:05:58,560 --> 00:05:59,800
有用户留言呢

147
00:05:59,800 --> 00:06:01,680
就说这个风格是猴来

148
00:06:01,680 --> 00:06:05,420
那我就让JPT美化一下

149
00:06:05,420 --> 00:06:06,480
当前的图片

150
00:06:06,480 --> 00:06:09,280
它最终优化成这个样子

151
00:06:09,280 --> 00:06:10,880
我认为它优化的

152
00:06:10,880 --> 00:06:12,240
还是非常不错的

153
00:06:12,240 --> 00:06:14,620
它是保留了原图的主题

154
00:06:14,620 --> 00:06:15,940
还有主要动作

155
00:06:15,940 --> 00:06:17,920
重新做了角色造型

156
00:06:17,920 --> 00:06:19,080
在让GPT呢

157
00:06:19,080 --> 00:06:19,980
生成一张图片

158
00:06:19,980 --> 00:06:22,720
要有常见的250字

159
00:06:22,720 --> 00:06:24,400
以及对应的拼音

160
00:06:24,400 --> 00:06:26,720
那可以看到

161
00:06:26,720 --> 00:06:27,780
有些字呢

162
00:06:27,780 --> 00:06:29,200
是出错了

163
00:06:29,200 --> 00:06:31,060
比如说右下角有两个五

164
00:06:31,060 --> 00:06:34,020
它检查过一次

165
00:06:34,020 --> 00:06:36,100
后来又修改过

166
00:06:36,100 --> 00:06:37,880
还是有问题

167
00:06:37,880 --> 00:06:40,260
那GBT在这里呢就翻车了

168
00:06:40,260 --> 00:06:42,800
接下来我让GBT做一个创意写作

169
00:06:42,800 --> 00:06:44,540
并且要输出长途

170
00:06:44,540 --> 00:06:46,760
它一第一版其实生成的也行

171
00:06:46,760 --> 00:06:50,600
但我觉得总是有点普通

172
00:06:50,600 --> 00:06:52,900
所以又让它优化了一下

173
00:06:52,900 --> 00:06:54,580
多加一些图片素材

174
00:06:54,580 --> 00:06:57,560
再让GBT呢做一个漫画风格

175
00:06:57,560 --> 00:06:59,980
来拆解洋涌的动作

176
00:06:59,980 --> 00:07:03,380
打腿 划水 露水 移壁

177
00:07:03,977 --> 00:07:12,057
擅長游泳的朋友們可以在彈幕裡評價一下GPT生成的效果如何,是否和真實的洋湧是一致的。

178
00:07:12,057 --> 00:07:16,217
当你在GPT里面让它帮你去做一个应用

179
00:07:16,217 --> 00:07:19,677
无论是web端或者是安卓苹果

180
00:07:19,677 --> 00:07:21,457
macOS应用的时候

181
00:07:21,457 --> 00:07:24,797
它总是先倾向于生成一张图片

182
00:07:24,797 --> 00:07:27,917
然后再根据图片来做设计

183
00:07:27,917 --> 00:07:33,477
我让2.5来生成三个理发web页面的预览图

184
00:07:33,477 --> 00:07:35,897
要给我三个不同的设计方向

185
00:07:35,897 --> 00:07:37,537
这是第一个

186
00:07:37,537 --> 00:07:38,817
第二个

187
00:07:38,817 --> 00:07:40,257
第三个

188
00:07:40,257 --> 00:07:42,237
总体还是非常不错的

189
00:07:42,237 --> 00:07:43,977
我最喜欢的就是第三个

190
00:07:43,977 --> 00:07:45,177
这样的一个设计风格

191
00:07:45,177 --> 00:07:47,157
OpenAI今天还发了一篇文章

192
00:07:47,157 --> 00:07:49,857
然后介绍他们的下一代模型

193
00:07:49,857 --> 00:07:51,057
有多么多么的厉害

194
00:07:51,057 --> 00:07:52,817
解决了一个什么问题

195
00:07:52,817 --> 00:07:55,577
那我就将这篇文章发给GBT

196
00:07:55,577 --> 00:07:57,857
让他来做一张

197
00:07:57,857 --> 00:08:00,057
很像的学术论文海报

198
00:08:00,057 --> 00:08:03,717
要包含来源中的重要图表

199
00:08:03,717 --> 00:08:05,157
图解和绘图

200
00:08:05,157 --> 00:08:08,197
最终他生成的是这样子的

201
00:08:08,197 --> 00:08:10,477
我个人觉得还是非常不错的

202
00:08:10,477 --> 00:08:11,737
接着再让它生成

203
00:08:11,737 --> 00:08:13,957
适合抹书的风格

204
00:08:13,957 --> 00:08:15,317
也是非常不错

205
00:08:15,317 --> 00:08:17,397
我将这样的图片发给它

206
00:08:17,397 --> 00:08:19,137
我说把它变成

207
00:08:19,137 --> 00:08:21,517
真实的电商的

208
00:08:21,517 --> 00:08:24,017
直播间手机截图页面

209
00:08:24,017 --> 00:08:25,517
展示所卖的扫帚

210
00:08:25,517 --> 00:08:27,337
并且要生成扫帚

211
00:08:27,337 --> 00:08:29,197
一系列的高质量商品图

212
00:08:29,197 --> 00:08:32,097
这是它生成的直播的页面

213
00:08:32,097 --> 00:08:33,817
和我原图

214
00:08:33,817 --> 00:08:36,297
相似度还是非常高

215
00:08:36,297 --> 00:08:38,177
这是它生成的商品图

216
00:08:38,177 --> 00:08:41,137
那我觉得它一开始生成的商品图里呢

217
00:08:41,137 --> 00:08:42,557
没有文字

218
00:08:42,557 --> 00:08:44,597
所以我就让它加上文字

219
00:08:44,597 --> 00:08:46,997
由于我选择的是Astro模型

220
00:08:46,997 --> 00:08:50,317
所以它的质地程度还是非常高的

221
00:08:50,317 --> 00:08:51,837
商品图上的文案呢

222
00:08:51,837 --> 00:08:53,137
还是非常高级的

223
00:08:53,137 --> 00:08:53,937
最后呢

224
00:08:53,937 --> 00:08:56,657
我就让它生成一张图片

225
00:08:56,657 --> 00:08:58,377
包含多张分镜

226
00:08:58,377 --> 00:08:59,697
然后调用MiniMax

227
00:08:59,697 --> 00:09:03,257
H3Max来生成视频

228
00:09:03,257 --> 00:09:05,337
这是它一开始生成的分镜

229
00:09:05,337 --> 00:09:06,437
深层的视频呢

230
00:09:06,437 --> 00:09:08,057
质感还是非常非常好的

231
00:09:08,057 --> 00:09:24,757
刚刚有提到GBT

232
00:09:24,757 --> 00:09:27,617
他在图片编辑的时候

233
00:09:27,617 --> 00:09:30,557
做的是精卷编辑

234
00:09:30,557 --> 00:09:32,057
那我这里呢

235
00:09:32,057 --> 00:09:34,377
是给他一张图提示GBT

236
00:09:34,377 --> 00:09:37,657
来换成五种不同风格的衣服

237
00:09:37,657 --> 00:09:38,977
要酷要有趣味

238
00:09:38,977 --> 00:09:40,377
花了五分钟

239
00:09:40,377 --> 00:09:41,577
它就帮我生成了

240
00:09:41,577 --> 00:09:43,957
当我切换左边不同图片的时候

241
00:09:43,957 --> 00:09:45,077
我们可以看到

242
00:09:45,077 --> 00:09:47,917
兔子的穿着在改变

243
00:09:47,917 --> 00:09:50,057
兔子的位置呢

244
00:09:50,057 --> 00:09:51,497
基本上没有在变动

245
00:09:51,497 --> 00:09:53,077
这还是非常符合

246
00:09:53,077 --> 00:09:55,257
OpenAI宣传的精准编辑的

247
00:09:55,257 --> 00:09:57,457
以上呢就是今天介绍

248
00:09:57,457 --> 00:09:59,817
XLGBT Image 2.5的所有内容

249
00:09:59,817 --> 00:10:03,637
那它可以帮我们做定格动画

250
00:10:03,637 --> 00:10:06,597
然后它的速度现在快了好多

251
00:10:06,597 --> 00:10:08,737
这在平常工作里

252
00:10:08,737 --> 00:10:11,557
真的是有一个非常大的帮助

253
00:10:11,557 --> 00:10:13,717
希望我的视频对你有帮助

254
00:10:13,717 --> 00:10:15,197
我们下次再见