WEBVTT 00:00.360 --> 00:03.960 -: Hey, I'm gonna walk you through how to prompt DALL-E 3. 00:03.960 --> 00:06.900 It's only available right now through GPT-4, 00:06.900 --> 00:10.320 which is on the paid plan, which you need to have, 00:10.320 --> 00:12.900 but it's pretty accessible once you do have that. 00:12.900 --> 00:15.150 So you can literally just ask it for an image 00:15.150 --> 00:17.100 and it's gonna come up with something. 00:17.100 --> 00:21.030 Give me an image of a corgi 00:21.030 --> 00:25.053 sitting on top of the Brandenburg Gate. 00:26.100 --> 00:31.020 And it's pretty useful for these types of queries 00:31.020 --> 00:32.880 where you're actually asking it 00:32.880 --> 00:35.820 to put something on top of something or behind something. 00:35.820 --> 00:39.660 This is something that the other AI image tools 00:39.660 --> 00:40.890 are currently not very good at. 00:40.890 --> 00:44.490 So Midjourney doesn't have that much situational awareness 00:44.490 --> 00:48.570 as well as Stable Diffusion also struggles with this. 00:48.570 --> 00:49.860 That said, the latest version 00:49.860 --> 00:52.980 of Stable Diffusion is a diffusion transformer model 00:52.980 --> 00:56.400 and it's rumored that DALL-E is the same. 00:56.400 --> 00:59.070 And what that means is it has an LLM 00:59.070 --> 01:02.070 that also powers the image generation. 01:02.070 --> 01:03.660 It's not just a diffusion model. 01:03.660 --> 01:05.160 It understands a little bit more 01:05.160 --> 01:06.690 about what's supposed to be in the image 01:06.690 --> 01:09.450 and it can reason a little bit better about that. 01:09.450 --> 01:10.590 So here we go. 01:10.590 --> 01:13.080 Now we've got this image, which is pretty cool. 01:13.080 --> 01:15.030 There's a few different things we can do here. 01:15.030 --> 01:17.130 So one is we can download the image. 01:17.130 --> 01:19.800 We could also click on the image. 01:19.800 --> 01:22.320 You can see that it's coming back here. 01:22.320 --> 01:25.770 Now we have some additional things that we can do. 01:25.770 --> 01:27.900 So we can click here to see the prompt 01:27.900 --> 01:30.030 and see what prompt it's written for us. 01:30.030 --> 01:32.790 This is one of the core things about accessibility, 01:32.790 --> 01:34.590 is that you don't need to know prompt engineering 01:34.590 --> 01:35.640 to use DALL-E. 01:35.640 --> 01:37.950 It's gonna write this for you here. 01:37.950 --> 01:41.580 You can see that it's expanded upon 01:41.580 --> 01:45.690 our initial just simple prompt to give us a much better one. 01:45.690 --> 01:48.600 But we're gonna say, first of all, we could just chat back 01:48.600 --> 01:52.560 and say the dog is supposed 01:52.560 --> 01:55.290 to look realistic. 01:55.290 --> 01:59.220 'Cause a lot of the times it's not, which is a real problem. 01:59.220 --> 02:01.020 And this is really useful 02:01.020 --> 02:04.140 because you can just talk in natural language and get back, 02:04.140 --> 02:06.030 it's gonna write another prompt for us. 02:06.030 --> 02:07.620 It's creating the image. 02:07.620 --> 02:09.510 So we're gonna get back something different. 02:09.510 --> 02:10.343 It's gonna respond, 02:10.343 --> 02:12.330 it's gonna write a different prompt for us, 02:12.330 --> 02:13.593 which is really helpful. 02:32.850 --> 02:33.683 Okay, here we go. 02:33.683 --> 02:36.570 Now we have a realistic dog, which is much better. 02:36.570 --> 02:37.870 Let's give it a thumbs up. 02:38.970 --> 02:41.160 And it describes what's going on. 02:41.160 --> 02:42.780 But the problem is we actually wanted it 02:42.780 --> 02:44.130 on top of the gate here. 02:44.130 --> 02:45.780 So I'm gonna click into the image. 02:45.780 --> 02:48.150 And one of the cool things they added more recently 02:48.150 --> 02:50.560 is you can select 02:51.993 --> 02:53.730 a certain part of it 02:53.730 --> 02:56.943 and you can get rid of that. 03:01.080 --> 03:01.990 And 03:05.610 --> 03:07.380 you can then prompt again. 03:07.380 --> 03:11.403 You can say Brandenburg, no dog. 03:12.630 --> 03:13.623 See what that does. 03:14.610 --> 03:18.390 So it should just get rid of this part here. 03:18.390 --> 03:19.790 So this is called inpainting 03:20.760 --> 03:22.860 and this used to be available at DALL-E 2. 03:22.860 --> 03:24.840 It's not available anymore unfortunately, 03:24.840 --> 03:28.593 but we can get access it in ChatGPT here. 03:44.280 --> 03:45.113 Okay, there we go. 03:45.113 --> 03:46.620 It's removed the dog from the image, 03:46.620 --> 03:48.210 which is really cool. 03:48.210 --> 03:52.680 Now we wanna click on this image, 03:52.680 --> 03:54.270 wanna potentially edit this 03:54.270 --> 03:58.143 and say, should say put the Corgi here, 04:00.000 --> 04:03.930 should say giant corgi 04:03.930 --> 04:07.053 on top of Brandenburg Gate. 04:11.220 --> 04:13.710 And again, this is just gonna change this part here. 04:13.710 --> 04:15.900 So if we're happy with the rest of the image, 04:15.900 --> 04:20.043 then this is a really useful way to improve things. 04:58.290 --> 04:59.123 Here we go. 04:59.123 --> 05:01.590 Now we have a Corgi on top of the Brandenburg Gate. 05:01.590 --> 05:03.660 So that's pretty helpful. 05:03.660 --> 05:06.690 And we're not gonna go too deep into this. 05:06.690 --> 05:09.600 The functionality is changing over time. 05:09.600 --> 05:13.380 But one thing I will mention though is that you can, 05:13.380 --> 05:17.880 if you go into here, you go into Explore GPTs, 05:17.880 --> 05:21.060 that you could find, the specific DALL-E one. 05:21.060 --> 05:24.330 You can use these where it only does images, 05:24.330 --> 05:26.820 and there's some that people have made that are custom. 05:26.820 --> 05:28.530 So this one's Cartoonize Yourself. 05:28.530 --> 05:30.540 It's got specific prompts for that. 05:30.540 --> 05:32.310 And you know, here's one 05:32.310 --> 05:34.683 that's really good at logos, et cetera. 05:36.840 --> 05:38.340 Here, this one actually is pretty cool. 05:38.340 --> 05:40.890 It creates prompts that are good for Midjourney 05:40.890 --> 05:42.330 so you can try that. 05:42.330 --> 05:43.200 All right, enjoy. 05:43.200 --> 05:45.210 Hopefully you have a good time with DALL-E 05:45.210 --> 05:47.550 and you can also access it via API, 05:47.550 --> 05:50.733 but primarily people are using it through the interface.