WEBVTT 00:00.880 --> 00:01.480 All right. 00:01.520 --> 00:03.240 Advanced consistent characters. 00:03.240 --> 00:08.480 So how do you take a character in one image and use the exact same character in a different scene? 00:08.480 --> 00:12.960 So this is the same guy in a business suit or playing tennis? 00:13.000 --> 00:14.480 I have a little trick for doing this. 00:14.480 --> 00:20.280 I used to do this in Midjourney, and I figured out a way to do it in foul recently, so that's what 00:20.280 --> 00:21.560 we're trying to accomplish. 00:21.560 --> 00:24.000 I'll just show you guys this how it works. 00:24.720 --> 00:26.520 We're going to run through it one by one. 00:27.000 --> 00:33.960 So first we need to just install the foul client, make sure it's there and then load the environment 00:33.960 --> 00:37.880 variables so we can get a API key for foul the foul key. 00:38.320 --> 00:40.640 So make sure you've done that there. 00:41.080 --> 00:44.160 And then we just need to bring in the foul client. 00:45.240 --> 00:50.920 So let's import foul client import request Pil image bytes IO. 00:51.520 --> 00:54.400 And then get the Python display as well. 00:55.480 --> 00:55.880 All right. 00:55.920 --> 01:04.680 And we want to take the original image which is this rhino in a suit. 01:05.650 --> 01:08.650 And then we want to shrink the image. 01:08.690 --> 01:09.090 Oh, actually. 01:09.090 --> 01:09.330 Sorry. 01:09.370 --> 01:15.050 We want to basically create a new image which will double the width. 01:15.090 --> 01:15.330 Right. 01:15.370 --> 01:22.410 So create a new image with double width and white background. 01:24.010 --> 01:24.290 All right. 01:24.330 --> 01:29.130 So this doubles the width and that keeps the original height. 01:29.530 --> 01:32.050 And then we're just creating this new image. 01:32.330 --> 01:36.210 And then we just paste the original image on the left hand side. 01:37.010 --> 01:43.490 And then we want to just save that and upload that to file. 01:45.090 --> 01:45.930 So we have that. 01:46.290 --> 01:55.690 And then here what we're doing is we are taking the the image size and basically creating a mask on 01:55.690 --> 01:56.610 the right hand side. 01:56.610 --> 01:59.170 So it's going to make the right hand side of the image black. 01:59.530 --> 02:02.810 And it's going to make the left hand side of the image white. 02:03.130 --> 02:12.450 And we basically just need to mask image paste paste that original image region on the canvas, and 02:12.450 --> 02:14.370 then we can upload that as Tim mask. 02:15.130 --> 02:16.050 Get the URL. 02:16.570 --> 02:18.610 And then we're going to see both these side by side. 02:18.610 --> 02:20.330 So this is all the hard work. 02:20.330 --> 02:24.530 Essentially it's just to create this new image with a section on the right. 02:24.530 --> 02:25.850 And you can't really see this very well. 02:25.850 --> 02:27.690 But this image goes all the way across here. 02:28.050 --> 02:35.250 So paste that image on here and then adds this section on the right hand side, which is a white section. 02:35.250 --> 02:37.210 And then it adds a black mask. 02:37.450 --> 02:39.650 So these two are going to be uploaded separately. 02:40.170 --> 02:43.690 And then from there on out it's basically just inpainting. 02:43.690 --> 02:52.210 And the reason why this works is that we have when you're inpainting, it takes into account the rest 02:52.210 --> 02:53.130 of the image. 02:53.410 --> 02:56.370 It'll stay consistent with that image. 02:56.370 --> 02:57.730 And that's why this works. 02:57.730 --> 03:01.650 Because if you just create something new from the prompt, it hasn't seen the other image. 03:01.850 --> 03:04.610 And even with image to image, it doesn't always work that well. 03:04.770 --> 03:08.610 Whereas when you're inpainting it has better prompt adherence is what I find. 03:08.650 --> 03:10.690 So let's see what that looks like. 03:10.890 --> 03:15.330 The prompt output is I've just again, we're just prompting for the right hand side of the image. 03:15.850 --> 03:20.130 We're feeding it in the left and we're saying on the right, like making it look like a caption. 03:20.170 --> 03:24.130 The photo of the same rhino wearing tennis clothes, carrying a tennis racket, playing tennis on a 03:24.130 --> 03:24.810 tennis court. 03:25.370 --> 03:28.090 And we're using this fill model, which is really good for out painting. 03:28.130 --> 03:29.730 I think that's super important. 03:29.770 --> 03:31.930 We've got the new width and the new height here. 03:31.930 --> 03:35.370 We're passing that in the image size and here we go. 03:35.610 --> 03:39.970 We've got the same rhino but playing tennis right. 03:40.010 --> 03:40.890 So really cool. 03:41.090 --> 03:42.850 And yeah you could do whatever you want. 03:43.010 --> 03:49.290 The really useful thing I found is if you wanted to fine tune on the same character, you could go and 03:49.290 --> 03:54.290 generate a bunch of these, you know, in different scenarios, different situations, and then chop 03:54.290 --> 04:00.770 them all up programmatically, divide the image in half, and then you've got a load of different images 04:00.770 --> 04:02.570 of the same character in different spaces. 04:02.610 --> 04:05.090 And then that could be your fine tuning data. 04:05.130 --> 04:07.210 So that's one of the ways I use this. 04:07.210 --> 04:09.370 But yeah, lots of different ways to tease. 04:09.370 --> 04:10.690 This is a good little hack. 04:10.890 --> 04:17.450 Again, I did it in Midjourney primarily initially, but now you can program and make that programmatic 04:17.450 --> 04:17.650 in.