WEBVTT 00:00.930 --> 00:02.820 -: All right, this is a really fun one, 00:02.820 --> 00:05.760 which is Gemini 2.0 image generation. 00:05.760 --> 00:07.350 And what makes this fun 00:07.350 --> 00:09.450 is not that it's an image generation model, 00:09.450 --> 00:11.340 which is neither here nor there. 00:11.340 --> 00:13.410 There's a lot of image generation models. 00:13.410 --> 00:17.700 But what's cool about this is that it has native editing. 00:17.700 --> 00:22.560 When you say, make me an image with the words hello Mike, 00:22.560 --> 00:25.050 it actually really understands the prompt intent 00:25.050 --> 00:26.400 as straight into an image. 00:26.400 --> 00:31.140 It is not actually calling an image generation model, 00:31.140 --> 00:32.610 like ChatGPT is. 00:32.610 --> 00:36.120 It's much better, I found, at prompt adherence. 00:36.120 --> 00:40.290 But a really cool thing that I think is the killer feature 00:40.290 --> 00:42.870 is that it can take an existing image, 00:42.870 --> 00:46.920 and because it understands a lot about what's in that image 00:46.920 --> 00:49.380 and it's not converting it into a prompt, 00:49.380 --> 00:52.140 then you can get some pretty cool stuff. 00:52.140 --> 00:53.850 I'm just gonna upload an image here. 00:53.850 --> 00:56.100 This is from a marketing website 00:56.100 --> 01:01.100 and this is specific '90s nostalgic gaming-style pixelated. 01:01.110 --> 01:06.110 So I'm just gonna say, create another image of this style, 01:07.320 --> 01:10.126 but it's a major city. 01:10.126 --> 01:13.680 (keyboard clacking) 01:13.680 --> 01:15.810 And if I do this in ChatGPT, 01:15.810 --> 01:17.430 it tends to get pretty bad results 01:17.430 --> 01:20.340 because it misunderstands the image style. 01:20.340 --> 01:24.000 But because we're not converting into a prompt 01:24.000 --> 01:26.490 and then converting to an image generation model, 01:26.490 --> 01:29.430 which might lose some of the information along the way, 01:29.430 --> 01:33.180 instead it's really bringing that same style across, 01:33.180 --> 01:34.170 which is very cool. 01:34.170 --> 01:35.490 And then, I'm just gonna say, 01:35.490 --> 01:40.490 make it of Times Square but in the same style. 01:40.532 --> 01:41.939 (keyboard clacking) 01:41.939 --> 01:42.772 Okay. 01:51.510 --> 01:53.430 There you go. Very cool. 01:53.430 --> 01:54.810 All right, I'm gonna try one more. 01:54.810 --> 01:57.773 And this is another really interesting way to do it. 01:57.773 --> 02:01.350 Because it understands what's in the image, 02:01.350 --> 02:03.990 it can also do character consistency. 02:03.990 --> 02:07.020 So I'm just gonna find a headshot. 02:07.020 --> 02:08.043 Let me find one. 02:11.040 --> 02:12.340 Yeah, there's my headshot. 02:14.580 --> 02:18.794 I'm gonna say, make a passport photo of this guy. 02:18.794 --> 02:21.794 (keyboard clacking) 02:26.010 --> 02:28.020 And there you go. (laughs) 02:28.020 --> 02:28.950 It's pretty good, 02:28.950 --> 02:31.320 considering that we didn't fine-tune a model here. 02:31.320 --> 02:33.720 This has been possible with fine-tuning for a while, 02:33.720 --> 02:36.720 but this is one-shot character consistency, 02:36.720 --> 02:38.580 which is pretty powerful. 02:38.580 --> 02:41.070 So yeah, I encourage you to mess around with it 02:41.070 --> 02:43.740 and think about how you wanna use it 02:43.740 --> 02:47.640 because there are many new applications you can create 02:47.640 --> 02:50.730 with this type of one-shot application, 02:50.730 --> 02:54.930 with this type of one-shot image-to-image type model 02:54.930 --> 02:56.940 or text-to-image type model. 02:56.940 --> 02:58.860 Because it's working natively, 02:58.860 --> 03:01.290 it can do so much more and it's really surprising 03:01.290 --> 03:04.890 how much more it can do without losing information 03:04.890 --> 03:07.080 in between translating into a prompt 03:07.080 --> 03:08.640 for an image generation model. 03:08.640 --> 03:10.050 It's all the same model now 03:10.050 --> 03:12.153 and it feels completely different.