WEBVTT 00:00.600 --> 00:02.883 -: Hey, let's talk about DALL-E 3. 00:03.750 --> 00:06.510 So DALL-E 3 is a diffusion model at its core 00:06.510 --> 00:08.970 and that means it can generate images 00:08.970 --> 00:12.480 from random noise given a prompt 00:12.480 --> 00:15.660 and the diffusion model works by first to take lots 00:15.660 --> 00:17.640 of images and add random noise to them 00:17.640 --> 00:19.830 and then see if the model kind of worked backwards. 00:19.830 --> 00:21.060 That's how it's trained. 00:21.060 --> 00:23.640 The DALL-E model is closed source, 00:23.640 --> 00:25.410 so we don't know exactly how it works, 00:25.410 --> 00:26.520 but some people conjecture 00:26.520 --> 00:29.580 it's also a transform model in there 00:29.580 --> 00:31.080 as well using the Diffusion model. 00:31.080 --> 00:33.630 So that's why it's so good at being able 00:33.630 --> 00:36.390 to place certain things in different parts of the image, 00:36.390 --> 00:39.180 unlike a lot of image generation models 00:39.180 --> 00:42.363 because it actually has some reason to. 00:43.740 --> 00:45.630 DALL-E 3 is created by OpenAI. 00:45.630 --> 00:47.790 They also make ChatGPT. 00:47.790 --> 00:49.590 It was released before Midjourney 00:49.590 --> 00:51.870 and Stable Diffusion, the first version 00:51.870 --> 00:53.130 and DALL-E 2 was the one 00:53.130 --> 00:56.430 that was first really good at image generation. 00:56.430 --> 00:58.950 Kinda kicked off that whole craze in 2022, 00:58.950 --> 01:02.760 but version three only came out in 2023. 01:02.760 --> 01:06.450 And then it's also available through ChatGPT 01:06.450 --> 01:08.760 and via the API. 01:08.760 --> 01:11.700 What are some key features of this compared 01:11.700 --> 01:13.740 to other image generation tools? 01:13.740 --> 01:15.480 One is that it can do in painting 01:15.480 --> 01:18.660 and they can do that directly in line in ChatGPT interface, 01:18.660 --> 01:19.650 which is really cool. 01:19.650 --> 01:21.540 So you can scrub out part of the image 01:21.540 --> 01:24.000 and then replace it with by prompting. 01:24.000 --> 01:25.290 It's really good at that actually, 01:25.290 --> 01:27.180 probably the best at that. 01:27.180 --> 01:30.150 It can also create prompts. 01:30.150 --> 01:31.800 This isn't specifically DALL-E 3, 01:31.800 --> 01:33.210 this is ChatGPT doing this, 01:33.210 --> 01:34.740 but ChatGPT can write a prompt 01:34.740 --> 01:38.160 for DALL-E 3 in the ChatGPT interface 01:38.160 --> 01:39.660 and you can see that prompt. 01:39.660 --> 01:43.020 So, you can click the I when you click into an image 01:43.020 --> 01:45.150 and it'll tell you what the prompt was 01:45.150 --> 01:46.530 and that's really helpful. 01:46.530 --> 01:49.290 The other thing is that you can just edit by talking. 01:49.290 --> 01:50.123 So you can say, 01:50.123 --> 01:51.990 "Oh no, I wanted this dog to look realistic." 01:51.990 --> 01:55.710 And it can change the way the images look cool. 01:55.710 --> 01:58.140 One of the problems is that there's no negative prompting. 01:58.140 --> 02:00.750 You can't really get rid of things in the image like you can 02:00.750 --> 02:02.820 with Stable Diffusion or Midjourney. 02:02.820 --> 02:05.460 There's no classifier free guidance. 02:05.460 --> 02:07.830 There's no real control at all on the parameters 02:07.830 --> 02:10.800 that you're using as you would get from some 02:10.800 --> 02:12.570 of the other platforms. 02:12.570 --> 02:13.770 There's no way to fine tune it 02:13.770 --> 02:15.720 like you can with Stable Diffusion. 02:15.720 --> 02:18.270 There's also a limited use of artist names. 02:18.270 --> 02:20.910 You can't use any artist in living memory 02:20.910 --> 02:22.560 like in the last 100 years. 02:22.560 --> 02:25.620 It's also a lot of content moderation, so you can't, 02:25.620 --> 02:28.020 for example, generate a picture of Donald Trump. 02:29.340 --> 02:31.230 Some of the use cases, given that some 02:31.230 --> 02:33.180 of the weaknesses turn into strength here. 02:33.180 --> 02:34.770 It's really great for children's stories 02:34.770 --> 02:38.880 'cause it never generates anything really bad or sensitive. 02:38.880 --> 02:39.990 So, that's helpful. 02:39.990 --> 02:41.310 It's good for blog images 02:41.310 --> 02:45.060 'cause it has like a cartoony vibe or style. 02:45.060 --> 02:48.030 And this is an example of a blog image that I generated. 02:48.030 --> 02:49.380 It's also quite fun for memes 02:49.380 --> 02:52.410 'cause it's quite good at text on images 02:52.410 --> 02:55.860 and it's quite good understanding concepts. 02:55.860 --> 02:57.750 I've been using it mostly for that. 02:57.750 --> 03:01.950 I don't use DALL-E 3 really that much in a serious capacity 03:01.950 --> 03:04.680 just because Stable Diffusion is much better 03:04.680 --> 03:06.330 to build your business on. 03:06.330 --> 03:09.000 It works quite well and it's open source. 03:09.000 --> 03:09.833 You don't have to worry 03:09.833 --> 03:11.520 about content moderation (indistinct). 03:11.520 --> 03:13.770 I would say DALL-E 3 is very good though 03:13.770 --> 03:16.653 and worth using for some use cases.