WEBVTT 00:00.150 --> 00:01.830 -: So you have a prompt and it's 00:01.830 --> 00:04.590 a pretty flowery prompt here. 00:04.590 --> 00:07.650 You have a lot of words and we're not 00:07.650 --> 00:09.930 a hundred percent sure whether we need all of these words. 00:09.930 --> 00:12.990 And obviously in Midjourney there's no real limit 00:12.990 --> 00:14.781 to how many tokens you can have, 00:14.781 --> 00:19.230 but it can, I think, add a lot of noise to the image, 00:19.230 --> 00:20.670 if you are pushing the model 00:20.670 --> 00:22.350 in too many different directions 00:22.350 --> 00:25.080 and maybe not all of these tokens are necessary, 00:25.080 --> 00:27.210 you could be getting random results 00:27.210 --> 00:28.987 or like unusual results because 00:28.987 --> 00:30.510 you're including a lot of things 00:30.510 --> 00:33.900 that aren't really contributing a lot, but adding noise. 00:33.900 --> 00:36.516 Okay, so I'm just gonna copy this 00:36.516 --> 00:39.243 and then I'm gonna say shorten, 00:40.224 --> 00:43.470 and this is a technique that I haven't actually seen 00:43.470 --> 00:46.590 that many people use, but it is pretty amazing. 00:46.590 --> 00:48.840 So if you just hit enter, I haven't seen 00:48.840 --> 00:51.870 any other model that has this. 00:51.870 --> 00:54.900 Maybe it'd be possible to build for stable diffusion, 00:54.900 --> 00:57.660 but what it does is it chooses 00:57.660 --> 01:00.302 what are the important tokens and it tells you 01:00.302 --> 01:02.430 here which ones are important. 01:02.430 --> 01:04.533 So graffiti style mural is really important. 01:04.533 --> 01:07.424 So clever is more important than slide. 01:07.424 --> 01:09.570 That's interesting. And then dachshund. 01:09.570 --> 01:11.758 It doesn't need dog showcasing Banksy. 01:11.758 --> 01:15.270 And then it's stencil. It doesn't need technique, right? 01:15.270 --> 01:16.830 It doesn't need set against. 01:16.830 --> 01:18.540 It just needs brick and background. 01:18.540 --> 01:21.195 It actually knows, because it is the model, 01:21.195 --> 01:24.765 it can give you the attention that's paid to each token 01:24.765 --> 01:27.782 and make a subjective decision about 01:27.782 --> 01:30.000 what should be included in the prompt. 01:30.000 --> 01:32.190 And then it gives you this pretty cool thing here, 01:32.190 --> 01:36.211 which is it just slowly narrows it down, deletes words. 01:36.211 --> 01:39.151 Then this is just deleting all the unnecessary words 01:39.151 --> 01:41.310 and then cuts more and more away 01:41.310 --> 01:43.774 until literally if you just put in like mural, 01:43.774 --> 01:46.440 dachshund, Banksy, brick, then 01:46.440 --> 01:47.940 you get something pretty good, 01:47.940 --> 01:50.550 even though you haven't used all of these tokens. 01:50.550 --> 01:52.320 So this is just an interesting concept. 01:52.320 --> 01:56.220 One thing I do a lot is I will use this feature 01:56.220 --> 01:59.430 and then copy the prompt across to stable diffusion. 01:59.430 --> 02:02.207 I know that's not quite exactly transferrable, 02:02.207 --> 02:05.070 but it is useful to know, at least in general, 02:05.070 --> 02:06.510 like what types of words are triggering 02:06.510 --> 02:08.340 for this type of model. 02:08.340 --> 02:10.034 All right. The other thing you can do is 02:10.034 --> 02:11.850 you can click on one of these and it will run that prompt. 02:11.850 --> 02:13.110 But there's one more little thing 02:13.110 --> 02:15.420 I wanna show you, which is show details. 02:15.420 --> 02:16.980 And this is really exciting because 02:16.980 --> 02:18.330 it gives you the numbers, right? 02:18.330 --> 02:20.130 It actually gives you the numbers. 02:20.130 --> 02:21.153 You can see there's a little chart here, 02:21.153 --> 02:23.880 the dachshund and is very important. 02:23.880 --> 02:25.950 Banksy and the mural, but then where as 02:25.950 --> 02:27.944 like showcasing as not as important. 02:27.944 --> 02:30.420 So that's how it makes these decisions. 02:30.420 --> 02:32.580 You can look and say, so I thought the word subtly, 02:32.580 --> 02:34.080 for example, would be important, 02:34.080 --> 02:35.490 but it's zero, point, zero, one 02:35.490 --> 02:37.470 in terms of attention paid to that token. 02:37.470 --> 02:39.900 So yeah, really not worth it. 02:39.900 --> 02:40.890 The other thing is it tells you 02:40.890 --> 02:42.240 a little bit how things are split up. 02:42.240 --> 02:46.260 So saying and social. So that is interesting 02:46.260 --> 02:48.510 because I've used the words stencil, art, 02:48.510 --> 02:50.580 technique, and social commentary. 02:50.580 --> 02:52.860 Like I thought that it would split. 02:52.860 --> 02:54.570 So it's like stencil, art, technique, 02:54.570 --> 02:56.490 and then social commentary would be separate, 02:56.490 --> 02:58.290 but it's actually paired and social together. 02:58.290 --> 03:01.710 So I might be getting unusual results because of that. 03:01.710 --> 03:04.320 Right, so it's really important to know that 03:04.320 --> 03:05.490 and to think about that when 03:05.490 --> 03:08.354 you are starting to split these prompts up 03:08.354 --> 03:10.335 and putting them in different sections. 03:10.335 --> 03:12.477 There's just really good intelligence. 03:12.477 --> 03:15.780 All right, so just out of interest, let's run number five 03:15.780 --> 03:18.640 and see what we get back, and hit submit 03:21.900 --> 03:24.360 and see if we get like a similar image, 03:24.360 --> 03:27.513 even though we've drastically shortened the prompt. 03:33.060 --> 03:34.800 While this is running out, I just wanna say as well, 03:34.800 --> 03:36.742 one thing this teaches me is just that 03:36.742 --> 03:40.170 it is not like a LLM, right? 03:40.170 --> 03:41.820 When you're prompting for images, 03:41.820 --> 03:44.730 it doesn't understand like the words necessarily. 03:44.730 --> 03:48.600 It's just grabbing the individual tokens more than anything. 03:48.600 --> 03:52.020 You might not actually be like using natural language 03:52.020 --> 03:54.030 to describe things like I did here where I said, 03:54.030 --> 03:57.360 and social commentary or a contrasted urban brick. 03:57.360 --> 03:58.890 You could literally just list words, 03:58.890 --> 04:01.320 I think, and get the similar results. 04:01.320 --> 04:04.175 Here we go. We actually got some pretty cool results here. 04:04.175 --> 04:06.510 Although it did get a little bit messed up here, 04:06.510 --> 04:07.920 and that's pretty amazing though. 04:07.920 --> 04:09.360 When you think we went from 04:09.360 --> 04:13.380 all these words down to just four key words. 04:13.380 --> 04:15.723 That's still a pretty amazing result.