WEBVTT 00:00.270 --> 00:02.430 -: All right, let's walk through RunwayML 00:02.430 --> 00:03.630 and the Gen-2 model, 00:03.630 --> 00:05.040 which is currently one of the best 00:05.040 --> 00:07.170 text-to-video models out there. 00:07.170 --> 00:11.070 And this is a diffusion model like Stable Diffusion. 00:11.070 --> 00:14.970 In fact, the RunwayML guys were part of the team 00:14.970 --> 00:17.220 that were behind the original Stable Diffusion. 00:17.220 --> 00:18.780 They focused more on video 00:18.780 --> 00:21.660 after that text-to-image model was released 00:21.660 --> 00:23.700 and been making a lot of strides. 00:23.700 --> 00:25.680 So pretty interesting stuff. 00:25.680 --> 00:29.010 The actual model itself is, 00:29.010 --> 00:31.890 you know, accessed via their web interface. 00:31.890 --> 00:35.190 They have a platform that you can use for video editing 00:35.190 --> 00:37.950 and there's a bunch of really interesting features in there, 00:37.950 --> 00:40.380 including some they've added more recently. 00:40.380 --> 00:43.680 The main features I think that people are excited about, 00:43.680 --> 00:45.540 the first feature really was text-to-video. 00:45.540 --> 00:49.800 So being able to type in a prompt and then get a video back. 00:49.800 --> 00:52.260 Now, bearing in mind this is pretty early days 00:52.260 --> 00:53.093 for text-to-video. 00:53.093 --> 00:55.800 You do get some, kind of, five second clips 00:55.800 --> 00:57.630 that might be close to usable, 00:57.630 --> 01:00.690 but, I would say in general it's not a great, 01:00.690 --> 01:03.120 it is not good enough quality 01:03.120 --> 01:06.810 to actually drop in directly into, you know, 01:06.810 --> 01:08.310 a feature film or something like that. 01:08.310 --> 01:11.457 This is more like B-Roll level quality, if that, 01:11.457 --> 01:13.530 and the performance is improving exponentially. 01:13.530 --> 01:16.320 So it's worth learning how it works now 01:16.320 --> 01:18.570 because in 6 months, 12 months, we'll start to see, 01:18.570 --> 01:22.470 I think movies, like actual movies or short stories 01:22.470 --> 01:24.390 being generated with this. 01:24.390 --> 01:28.410 Image-to-video is another thing they added more recently, 01:28.410 --> 01:30.300 which has been really powerful. 01:30.300 --> 01:33.360 And this allows you to give it a base image 01:33.360 --> 01:34.650 and then turn that into a video. 01:34.650 --> 01:36.150 That tends to work a lot better, 01:36.150 --> 01:38.010 particularly if you're using a really good image model, 01:38.010 --> 01:40.157 like Midjourney as the base image 01:40.157 --> 01:42.780 because Midjourney has really great detail, 01:42.780 --> 01:46.410 it's good for characters and then you can generate 01:46.410 --> 01:50.280 from the generate useful video, which is really helpful. 01:50.280 --> 01:51.810 And there's also this really cool feature 01:51.810 --> 01:53.700 that just brought out called Motion Brush, 01:53.700 --> 01:55.740 which is like inpainting for videos. 01:55.740 --> 01:58.080 And they do actually have inpainting for videos as well. 01:58.080 --> 02:00.660 But essentially this is letting you choose 02:00.660 --> 02:04.230 one specific part of the image and animate that image, 02:04.230 --> 02:05.640 that part of the image specifically, 02:05.640 --> 02:08.130 bucket that under inpainting, Motion Brush, 02:08.130 --> 02:09.240 similar kind of concept. 02:09.240 --> 02:12.360 Take an existing video and then change some parts 02:12.360 --> 02:14.430 of that video as well as they have, 02:14.430 --> 02:15.780 you can remove the background, 02:15.780 --> 02:18.330 you can change a specific aspect of the video. 02:18.330 --> 02:21.660 So that's, I think, the future is taking a lot 02:21.660 --> 02:23.130 of existing video footage 02:23.130 --> 02:24.980 and changing some elements of it, 02:24.980 --> 02:26.010 of video editing, 02:26.010 --> 02:28.890 I think is gonna be more useful more immediately 02:28.890 --> 02:32.433 than the purely generating a whole video from the prompt. 02:33.510 --> 02:36.420 And now what are the use cases for RunwayML right now? 02:36.420 --> 02:39.180 I would say creating B-Roll like this, 02:39.180 --> 02:40.770 give here this example. 02:40.770 --> 02:43.320 This is literally just me texting, 02:43.320 --> 02:45.390 you know, text prompting, saying, 02:45.390 --> 02:48.480 generate some aerial drone footage of Northern Ireland. 02:48.480 --> 02:50.700 It made up this house, made up these cars, et cetera. 02:50.700 --> 02:53.940 And you can still tell it's AI 02:53.940 --> 02:55.530 and it's particularly the text 02:55.530 --> 02:57.990 and the fact that the car is too long. 02:57.990 --> 03:00.540 But this is pretty great actually for B-Roll, 03:00.540 --> 03:03.510 especially if you're not producing like a Hollywood movie. 03:03.510 --> 03:05.550 If this is something your kind of, 03:05.550 --> 03:09.240 companies, like your startups, video production, 03:09.240 --> 03:11.040 and you're just creating like an explainer video 03:11.040 --> 03:13.080 or something like that, or you know, 03:13.080 --> 03:15.120 like a non-professional video, 03:15.120 --> 03:16.350 this can be pretty good. 03:16.350 --> 03:18.540 You can also do video editing here 03:18.540 --> 03:20.220 and you can see actually 03:20.220 --> 03:22.740 how well it picks out different aspects from the screen. 03:22.740 --> 03:25.320 So, all I did here is just click on, 03:25.320 --> 03:26.430 this is an existing video 03:26.430 --> 03:28.410 of a woman running through a field. 03:28.410 --> 03:30.060 I clicked, just clicked once on the woman 03:30.060 --> 03:31.980 and it found her exact outline. 03:31.980 --> 03:34.770 So I think it's pretty good for that actually. 03:34.770 --> 03:37.170 And you can remove a lot of parts of the image 03:37.170 --> 03:38.940 or kind of add different parts, 03:38.940 --> 03:40.020 erase the background, 03:40.020 --> 03:41.400 whatever it is you need. 03:41.400 --> 03:45.120 So I think that's gonna be like the main use case initially. 03:45.120 --> 03:48.270 That is RunwayML is just a brief introduction. 03:48.270 --> 03:52.620 Obviously there's a lot that goes into AI video generation 03:52.620 --> 03:54.660 and it's still pretty nascent, 03:54.660 --> 03:57.930 so I don't expect usable results straight away. 03:57.930 --> 03:59.280 But for those who are interested, 03:59.280 --> 04:01.620 hopefully this gives you a good understanding 04:01.620 --> 04:04.623 of generally what is being done with it right now.