WEBVTT 00:01.080 --> 00:04.680 -: Okay, we're gonna learn how to make consistent characters. 00:04.680 --> 00:08.070 The way this works is using in painting, 00:08.070 --> 00:11.940 and this is just a prompt that I've done earlier, 00:11.940 --> 00:13.050 I'm just gonna paste that in 00:13.050 --> 00:14.880 and I'll explain to you what we're doing. 00:14.880 --> 00:18.360 Consistent characters, you want to be able to 00:18.360 --> 00:21.900 have the model, kind of use the same person again 00:21.900 --> 00:23.460 and again in different situations. 00:23.460 --> 00:26.100 When you're creating a character for the book or a novel 00:26.100 --> 00:28.110 or a game, whatever it is, 00:28.110 --> 00:29.520 you're creating a character for, 00:29.520 --> 00:31.290 you want it to be consistent, you don't want it 00:31.290 --> 00:33.720 to hallucinate, look a little bit different. 00:33.720 --> 00:36.810 In each case, the trick is to ask it 00:36.810 --> 00:41.280 to generate two images side by side of a character 00:41.280 --> 00:45.870 and then, and then it's going to have two different frames, 00:45.870 --> 00:47.940 which will be really helpful 00:47.940 --> 00:50.790 because you've been able to remove one of those frames 00:50.790 --> 00:54.600 and then use in painting to fill in different poses. 00:54.600 --> 00:56.370 So I'm gonna show you how that works. 00:56.370 --> 00:57.734 But the, the key is 00:57.734 --> 01:02.100 that you put in into your prompt two images side by side, 01:02.100 --> 01:03.960 and then you put your prompt 01:03.960 --> 01:05.760 and then you typically, 01:05.760 --> 01:07.860 it's helpful if you say like photo booth portrait 01:07.860 --> 01:09.120 or something like that, just 01:09.120 --> 01:10.980 so it gets in a consistent format of 01:10.980 --> 01:12.720 where you can really see their face. 01:12.720 --> 01:15.300 And then the other thing here is we're setting the aspect 01:15.300 --> 01:17.130 ratio of two to one. 01:17.130 --> 01:21.060 And that means instead of the say 5 12 by 5 12 image, 01:21.060 --> 01:23.550 you're gonna get basically enough space for two of them. 01:23.550 --> 01:25.830 And that's gonna be helpful as well. 01:25.830 --> 01:28.233 So just waiting for that to run. 01:32.310 --> 01:35.700 Okay. And we're just gonna choose the one that we like. 01:35.700 --> 01:40.470 I think this guy at the top here looks like pretty cool. 01:40.470 --> 01:41.820 This one looks a little bit AI. 01:41.820 --> 01:45.483 I'm going to choose the top one. Gonna upscale that. 01:49.080 --> 01:50.520 And you can see it's the same person, 01:50.520 --> 01:52.920 it's just two different shots of the same person. 01:52.920 --> 01:55.410 One of his hair is kind of down a little bit 01:55.410 --> 01:56.910 and the shadow is a little bit different. 01:56.910 --> 01:59.160 You already have now two images of the same face, 01:59.160 --> 02:02.190 which is helpful, but you to really get the effect here, 02:02.190 --> 02:03.740 you can click on Vary (Region). 02:06.810 --> 02:08.700 And this is also possible in any other model 02:08.700 --> 02:10.890 that supports in painting by the way. 02:10.890 --> 02:12.090 So you're gonna go Vary (Region) 02:12.090 --> 02:13.780 and then you're just gonna select 02:14.910 --> 02:16.473 this whole area here, 02:18.240 --> 02:19.830 to erase. 02:19.830 --> 02:22.650 And then typically what you want to do is just kind 02:22.650 --> 02:25.260 of change a little bit of the prompt, right? 02:25.260 --> 02:28.680 Just gonna put this in here. 02:28.680 --> 02:30.150 I'm just gonna change this instead of 02:30.150 --> 02:31.890 two images side by side, 02:31.890 --> 02:35.620 I'm gonna change it to side profile image 02:37.612 --> 02:40.530 and then this is, everything else is the same, right? 02:40.530 --> 02:42.600 So we haven't changed anything else. 02:42.600 --> 02:44.550 So we'll see what that comes back with. 02:45.930 --> 02:49.500 And so that prompt is just gonna be for the, 02:49.500 --> 02:51.003 just the right hand side. 02:55.530 --> 02:58.527 And because we're using, in painting it, it takes the 02:58.527 --> 03:00.720 left hand photo as an input. 03:00.720 --> 03:02.727 So it will make the same person 03:02.727 --> 03:04.110 and it's the same prompt as well. 03:04.110 --> 03:07.560 So it's doubling up on consistency here. And there you go. 03:07.560 --> 03:09.093 You can see it diffusing in, 03:11.430 --> 03:13.883 but you can see that it's the same person as well. 03:23.940 --> 03:25.530 Now while that's waiting, I just want 03:25.530 --> 03:27.210 to explain what you could do with this. 03:27.210 --> 03:29.820 So one is you could just do this manually, right? 03:29.820 --> 03:31.710 Like any type of shot you need, 03:31.710 --> 03:33.180 if you need a zoomed out shot 03:33.180 --> 03:36.690 or you need jumping from a building, whatever it is you need 03:36.690 --> 03:38.820 for your creation, if you're making a novel 03:38.820 --> 03:40.620 or some sort of game 03:40.620 --> 03:43.710 or whatever it is, you could manually make these, right? 03:43.710 --> 03:46.740 And, and then as long as you have one frame consistent, 03:46.740 --> 03:49.577 you can keep making multiple frames and then up scaling them 03:49.577 --> 03:51.330 and using what you have 03:51.330 --> 03:53.700 and chopping them up into different pieces. 03:53.700 --> 03:57.390 But the, the other thing you can do here is if you generate 03:57.390 --> 03:59.640 enough of these images, you could then slice them 03:59.640 --> 04:00.900 all, in Photoshop. 04:00.900 --> 04:02.880 So you could, you know, chop them down the middle. 04:02.880 --> 04:06.960 Now you have 2, 5, 12 by five 12 images 04:06.960 --> 04:09.690 and you can use those to train a custom model. 04:09.690 --> 04:13.230 If you have done the lesson on DreamBooth 04:13.230 --> 04:15.090 or you understand how that works 04:15.090 --> 04:18.210 and Stable Diffusion, you could actually now train a model 04:18.210 --> 04:19.620 based on this character 04:19.620 --> 04:21.120 and be able to replicate this character. 04:21.120 --> 04:24.150 Because all it takes is between 10 and 30 images 04:24.150 --> 04:26.490 of someone for it to understand the, 04:26.490 --> 04:28.230 the character and understand the face. 04:28.230 --> 04:30.630 That could be really powerful because then you would have a 04:30.630 --> 04:32.220 custom model, you could then put 04:32.220 --> 04:33.810 that character into any situation 04:33.810 --> 04:35.160 and it would be consistent. 04:35.160 --> 04:36.300 This is a really helpful trick 04:36.300 --> 04:38.010 and once you understand it, 04:38.010 --> 04:41.130 I think it's also a technique you can use in many other 04:41.130 --> 04:43.050 situations where you need consistency to. 04:43.050 --> 04:44.742 So if you need a consistent building 04:44.742 --> 04:47.040 or you need a consistent backdrop 04:47.040 --> 04:50.130 or whatever it is, you can always just keep one half the 04:50.130 --> 04:51.930 image as the original 04:51.930 --> 04:54.720 and then only in paint the other half of the image 04:54.720 --> 04:56.610 to maintain that visual consistency. 04:56.610 --> 04:58.470 So yeah, this is a really big unlock 04:58.470 --> 04:59.970 for me and I use it all the time. 04:59.970 --> 05:02.120 Hopefully you guys will find it useful too.