WEBVTT 00:00.690 --> 00:02.400 -: All right, Stable Diffusion. 00:02.400 --> 00:04.170 What is it good at, what is it bad at? 00:04.170 --> 00:07.350 So the really interesting thing about stable 00:07.350 --> 00:08.820 diffusion is that it's open source. 00:08.820 --> 00:10.980 So I think that that's been a real benefit. 00:10.980 --> 00:13.140 A lot of people have innovated on top of it 00:13.140 --> 00:15.360 and figured out interesting ways to use it. 00:15.360 --> 00:19.890 So you do have to use a GPU graphics card to run it, 00:19.890 --> 00:22.800 but you can actually do that in the cloud, 00:22.800 --> 00:24.150 which is pretty cheap. 00:24.150 --> 00:27.810 Or for free, actually, in Google Colab. 00:27.810 --> 00:31.080 So it's a great way to learn, or if you have an M1 Mac 00:31.080 --> 00:33.480 or M2 Mac, that you can do that as well. 00:33.480 --> 00:35.460 You can use negative prompts, which is helpful. 00:35.460 --> 00:38.190 I think it gives you a bit more control flexibility, 00:38.190 --> 00:40.110 so you can kind of, in this case, 00:40.110 --> 00:42.180 we've got a photograph of astronaut riding a horse, 00:42.180 --> 00:44.580 but we don't want 'em to be in space. 00:44.580 --> 00:47.100 There's also classifier free guidance 00:47.100 --> 00:49.550 that kind of chooses how much the prompt matches, 00:50.880 --> 00:52.320 how much the image matches the prompt, 00:52.320 --> 00:55.020 how creative the AI can be, essentially. 00:55.020 --> 00:57.030 But the really cool thing, I think is Dreambooth. 00:57.030 --> 01:01.020 So that's the ability to train the AI on some concept 01:01.020 --> 01:03.420 or person, like this is a picture of me, 01:03.420 --> 01:05.670 but generated by AI. 01:05.670 --> 01:07.500 I think this is like the most advanced, 01:07.500 --> 01:10.920 like futuristic thing that AI can do at the minute. 01:10.920 --> 01:12.570 It's really cool. 01:12.570 --> 01:14.460 There's some limitations though. 01:14.460 --> 01:16.800 I think technical ability is a big one. 01:16.800 --> 01:18.630 You really have to know how to code 01:18.630 --> 01:19.590 to mess around with this. 01:19.590 --> 01:23.190 I think otherwise you're gonna get confused pretty quickly. 01:23.190 --> 01:24.990 It also takes like a long time to run. 01:24.990 --> 01:27.450 So unless you have powerful computer, 01:27.450 --> 01:29.340 it's gonna be like really difficult for you 01:29.340 --> 01:31.953 to iterate or kind of change anything about it. 01:32.790 --> 01:35.610 Also, runs into problems with faces and hands. 01:35.610 --> 01:37.920 So here's a few generations I've made 01:37.920 --> 01:39.900 where I've run into issues, 01:39.900 --> 01:42.060 although I think fine tuning the model, 01:42.060 --> 01:44.790 like using Dreambooth has definitely helped. 01:44.790 --> 01:46.440 So there's some people have made custom models 01:46.440 --> 01:48.940 that are really good at stock photos, for example. 01:49.800 --> 01:51.660 There's also kind of questionable provenance. 01:51.660 --> 01:54.600 So it's like, initially, it kind of came out 01:54.600 --> 01:56.460 as part of Stability AI, 01:56.460 --> 01:59.823 but then it was actually built by Runway, 02:01.620 --> 02:03.210 which is another AI company. 02:03.210 --> 02:05.790 But it was kind of funded by Stability AI. 02:05.790 --> 02:09.270 But then it was also a team from from Germany, 02:09.270 --> 02:10.650 the LMU Munich. 02:10.650 --> 02:14.160 So I think there were actually some issues 02:14.160 --> 02:16.140 where the model got taken down 02:16.140 --> 02:17.670 from the face for a while 02:17.670 --> 02:19.860 because it was fighting IP breach. 02:19.860 --> 02:22.140 But yeah, I mean, I think it's probably all sorted out 02:22.140 --> 02:26.880 now, but definitely, it's not as clear 02:26.880 --> 02:28.320 in terms of who owns it 02:28.320 --> 02:30.963 and who's responsible for improving it. 02:32.130 --> 02:34.200 But otherwise, I think Stable Diffusion 02:34.200 --> 02:37.290 is really the only AI art tool 02:37.290 --> 02:38.880 that you could really build your business on. 02:38.880 --> 02:41.100 Because it's open source, 02:41.100 --> 02:44.910 you don't need to worry about them blocking anything, right? 02:44.910 --> 02:47.070 Like you can run it on your own computer 02:47.070 --> 02:49.890 and then you can modify it the way that you want. 02:49.890 --> 02:51.450 You can fine tune it with Dreambooth. 02:51.450 --> 02:55.920 So I think that this has got the most commercial value out 02:55.920 --> 02:57.633 of any of the AI models.