WEBVTT

00:00.630 --> 00:02.040
-: All right, I'm gonna talk about how

00:02.040 --> 00:06.000
to get realistic models that are vaguely familiar

00:06.000 --> 00:08.700
and also attractive with Midjourney.

00:08.700 --> 00:10.800
So, this is a little trick I found.

00:10.800 --> 00:11.940
Someone, this isn't mine.

00:11.940 --> 00:14.940
Someone taught me this, but I can't remember who.

00:14.940 --> 00:17.190
If you are that person, you should remind me

00:17.190 --> 00:18.720
and I'll give you attribution.

00:18.720 --> 00:19.653
But one of the problems

00:19.653 --> 00:22.652
that you have in general is you say a photo of a person

00:22.652 --> 00:25.788
actually does generate pretty good stuff.

00:25.788 --> 00:28.140
I think Midjourney in particular is really good at

00:28.140 --> 00:30.450
generating a diverse set of images.

00:30.450 --> 00:33.540
But I have a little trick, which is you want the models

00:33.540 --> 00:35.340
to be, I think, attractive,

00:35.340 --> 00:38.010
but also you don't wanna say a photo

00:38.010 --> 00:40.470
of Emma Watson or whatever.

00:40.470 --> 00:43.020
Because, say if you want to use that in a,

00:43.020 --> 00:48.020
in context, you say a photo of Emma Watson as a pirate.

00:48.330 --> 00:51.180
If you want a picture of a pirate like Emma Watson is a good

00:51.180 --> 00:54.810
way to make that person like Hollywood attractive.

00:54.810 --> 00:56.661
But that can be a real problem

00:56.661 --> 01:00.390
because that is potentially violating her image rights

01:00.390 --> 01:03.060
and you can't really use her likeness.

01:03.060 --> 01:05.460
But there is a trick to doing this,

01:05.460 --> 01:06.630
and so we're just gonna wait

01:06.630 --> 01:08.280
for these generations to happen.

01:08.280 --> 01:09.720
The model, by the way, is very good at

01:09.720 --> 01:11.430
generating pictures of Emma Watson.

01:11.430 --> 01:15.270
This is why a lot of AI generation tutorials use her

01:15.270 --> 01:19.170
as an example because it has a lot of training data on her.

01:19.170 --> 01:20.130
And that's her as a pirate.

01:20.130 --> 01:23.040
And it looks, I think, a little bit too close to her.

01:23.040 --> 01:24.840
Although Midjourney does do some stuff,

01:24.840 --> 01:27.810
I think in the backend to change your prompt,

01:27.810 --> 01:30.000
so you're not getting an exact replica, which is,

01:30.000 --> 01:31.785
I think smart on their behalf.

01:31.785 --> 01:33.558
But here's my trick and

01:33.558 --> 01:36.150
what I'm gonna do is I'm gonna use prompt waiting

01:36.150 --> 01:38.400
and I'm gonna mix a couple of celebrities.

01:38.400 --> 01:40.170
This is an example here.

01:40.170 --> 01:45.170
I'm gonna, I'm gonna say a, a photo of, yeah,

01:45.360 --> 01:48.750
as a pirate and add this in here.

01:48.750 --> 01:50.520
And then, so what I'm mixing,

01:50.520 --> 01:52.860
and I, I think this actually works pretty well if you mix

01:52.860 --> 01:56.160
two different genders as well, a photo of Matthew Perry

01:56.160 --> 01:57.960
with 20%

01:57.960 --> 02:02.010
and a photo of Jennifer Lawrence with 80% as a pirate.

02:02.010 --> 02:05.160
So lemme just see what that comes back with.

02:05.160 --> 02:07.920
So I'm using a real mix of two different celebrities

02:07.920 --> 02:10.290
and what you get tend, what we tend

02:10.290 --> 02:12.750
to get at least, is when you mix two celebrities,

02:12.750 --> 02:13.950
they look vaguely familiar

02:13.950 --> 02:15.930
because you recognize some of the features,

02:15.930 --> 02:19.050
but it's not attributable to either celebrity.

02:19.050 --> 02:23.040
So, while that's running, I'm gonna do the opposite as well,

02:23.040 --> 02:24.540
just so you can compare.

02:24.540 --> 02:28.020
And now we're gonna dial it to 80%, Matthew Perry

02:28.020 --> 02:32.553
and 20%, Jennifer Lawrence as a pirate, and see what we get.

02:41.490 --> 02:43.230
Okay, there we go. So that is like,

02:43.230 --> 02:45.300
that picture in particular I think is pretty interesting

02:45.300 --> 02:47.400
because it looks a bit like Jennifer Lawrence,

02:47.400 --> 02:49.620
but is not really attributable to her,

02:49.620 --> 02:52.590
and then equally here, this is interesting.

02:52.590 --> 02:54.120
This is, you know, gonna be more male

02:54.120 --> 02:56.640
because we have more Matthew Perry in there,

02:56.640 --> 02:58.140
but it's very different.

02:58.140 --> 03:00.480
We could, we could try this as well with Emma Watson just

03:00.480 --> 03:02.610
to compare their results.

03:02.610 --> 03:06.557
And you can say a video of Matthew Perry, 20%

03:06.557 --> 03:11.557
and a photo of a Watson as a pirate.

03:12.090 --> 03:13.920
So you can mix those two things together

03:13.920 --> 03:15.360
and look at, this is pretty cool.

03:15.360 --> 03:19.020
I say that these images in particular, I think are like,

03:19.020 --> 03:20.580
they look like they should be a Hollywood

03:20.580 --> 03:22.350
actor, but they're not.

03:22.350 --> 03:25.590
They're fake. And you can get some really cool results.

03:25.590 --> 03:28.560
The other thing you can do is not just mixed genders,

03:28.560 --> 03:33.090
but just have two different people of the same gender,

03:33.090 --> 03:36.210
but like Amber heard in here, give a dash of Amber Heard

03:36.210 --> 03:38.550
to Emma Watson and see what it looks like.

03:38.550 --> 03:40.770
This one interestingly looks a little bit like the

03:40.770 --> 03:42.180
girl from Game of Thrones.

03:42.180 --> 03:44.850
So yeah. Anyway, this is, this is a technique

03:44.850 --> 03:46.380
that I use all the time.

03:46.380 --> 03:48.780
I very rarely prompt for a specific celebrity.

03:48.780 --> 03:52.320
I mix 'em together and you get a much better result,

03:52.320 --> 03:55.920
I think, and it's more unique, more non-attributable.

03:55.920 --> 03:57.930
So, really helpful technique.

03:57.930 --> 03:59.280
And once you see it,

03:59.280 --> 04:01.701
you'll be using it everywhere hopefully.

04:01.701 --> 04:03.600
And here we go. There's a photo

04:03.600 --> 04:05.250
of Amber Heard mixed with Emma Watson.

04:05.250 --> 04:06.810
I think it's still looks Hollywood,

04:06.810 --> 04:11.010
but it's, you can't tell which celebrity it is.

04:11.010 --> 04:12.760
All right. Hopefully that's useful.
