WEBVTT 00:00.270 --> 00:02.340 -: All right, so you're running Stable Diffusion. 00:02.340 --> 00:06.420 You've been using 1.5 probably because it's the best version 00:06.420 --> 00:09.120 and version two kind of didn't get great results. 00:09.120 --> 00:11.370 But now you're curious because you're hearing 00:11.370 --> 00:15.090 that people are getting good results with SDXL, the new one. 00:15.090 --> 00:16.800 It is actually very good 00:16.800 --> 00:19.140 and I'm gonna show you how to migrate to that. 00:19.140 --> 00:20.890 Here I have an image of 00:20.890 --> 00:24.750 an anime cat girl, classic Stable Diffusion, 00:24.750 --> 00:28.680 and I'm just gonna save the seed here. 00:28.680 --> 00:31.110 And then we've done 30 sampling steps 00:31.110 --> 00:34.230 and this was a Eular A sampling method. 00:34.230 --> 00:37.020 And we're gonna run the exact same prompt but with SDXL. 00:37.020 --> 00:38.970 So let's get it set up. 00:38.970 --> 00:42.900 So the first thing you need to do is go to the link. 00:42.900 --> 00:45.270 It's in Stability AI on Hugging Face. 00:45.270 --> 00:48.570 You need to download the base model, which is at the bottom. 00:48.570 --> 00:52.530 It's SDXL base 1.0 safetensors. 00:52.530 --> 00:54.210 And then you also need the refiner. 00:54.210 --> 00:57.153 So the way that this works is a little bit different. 00:58.080 --> 01:02.100 The refiner is optional, but it's worth trying out. 01:02.100 --> 01:05.970 In the files and versions, if you are SDXL refiner 1.0 01:05.970 --> 01:08.358 and these are safetensors files 01:08.358 --> 01:11.190 rather than the normal checkpoint files 01:11.190 --> 01:13.080 that you might be used to. 01:13.080 --> 01:15.909 Once you've downloaded them, they do take some time, 01:15.909 --> 01:18.630 so I'm not going to right now, I already have them. 01:18.630 --> 01:20.970 If you go to Stable Diffusion Web UI 01:20.970 --> 01:22.440 and then you go to models. 01:22.440 --> 01:25.530 And then in the models you should have a folder, 01:25.530 --> 01:27.570 which is Stable Diffusion. 01:27.570 --> 01:28.830 You stick 'em in there. 01:28.830 --> 01:30.630 And once they're in there, then they'll show up. 01:30.630 --> 01:33.540 And just also make sure you're on the latest version 01:33.540 --> 01:35.588 of automatic 1111. 01:35.588 --> 01:37.815 Cool, we have that working. 01:37.815 --> 01:41.970 This is version 1.5, let's upgrade to SDXL. 01:41.970 --> 01:46.110 So first thing we need to do is in my terminal, 01:46.110 --> 01:49.803 I'm just gonna Ctrl + C to get rid of that, stop it running. 01:51.450 --> 01:54.528 And normally I run bash webui.sh. 01:54.528 --> 01:57.930 But you want to run it slightly different, 01:57.930 --> 02:01.050 which is no-half-vae. 02:01.050 --> 02:03.480 You need that at least on my Mac. 02:03.480 --> 02:06.510 And obviously these instructions change fairly frequently. 02:06.510 --> 02:09.900 So just look up on the repository for automatic 1111, 02:09.900 --> 02:11.411 you should have a good tutorial, 02:11.411 --> 02:14.470 but that's how I got it working on my Mac. 02:14.470 --> 02:17.310 And it takes a little bit longer to set up. 02:17.310 --> 02:20.281 It is, I think a bit, obviously a bigger model. 02:20.281 --> 02:25.173 So we'll just see how this works. 02:27.000 --> 02:28.950 Pops open here. 02:28.950 --> 02:31.893 So then we're gonna copy across our prompt. 02:33.930 --> 02:38.930 So that one, then this is the negative prompt. 02:39.360 --> 02:42.840 We're gonna do 30 checkpoints. 02:42.840 --> 02:44.940 And then there's one thing we have to change here. 02:44.940 --> 02:49.940 Instead of 512, we need to make it 1024 by 1024. 02:50.490 --> 02:53.790 I just have a preset saved here using an extension. 02:53.790 --> 02:55.410 But you can just type that in. 02:55.410 --> 02:59.310 And then we also need to make sure our seed is in here, 02:59.310 --> 03:00.600 so copy the seed across. 03:00.600 --> 03:02.610 And this means that all things being equal, 03:02.610 --> 03:04.950 it would be the exact same image, 03:04.950 --> 03:06.210 but obviously things aren't equal 03:06.210 --> 03:09.540 because it's Stable Diffusion XL. 03:09.540 --> 03:12.990 That's everything we need to change apart from the refiner. 03:12.990 --> 03:15.070 The refiner model, where that comes in 03:15.910 --> 03:17.760 is if you click in here and choose refiner 03:17.760 --> 03:20.070 and you need to up here, actually, sorry, 03:20.070 --> 03:22.140 change this to the base. 03:22.140 --> 03:25.530 What the refiner does is the base will actually generate 03:25.530 --> 03:27.330 the composition of the image, 03:27.330 --> 03:29.090 but then the refiner, it switches over 03:29.090 --> 03:32.310 to the refiner at some point to add the fine details. 03:32.310 --> 03:34.800 So they've divided and conquered there 03:34.800 --> 03:36.030 with two different models. 03:36.030 --> 03:38.730 And one's really good at generating the base image 03:38.730 --> 03:40.920 and the other is really good at refining the image, 03:40.920 --> 03:42.450 hence the names. 03:42.450 --> 03:44.360 And we're also gonna set this to 0.6. 03:45.660 --> 03:47.250 You can experiment with this, 03:47.250 --> 03:48.780 but basically what this means is 03:48.780 --> 03:52.140 in the 30 sampling steps, when you get 60% of the way 03:52.140 --> 03:54.750 through, it's gonna switch over to the refiner model. 03:54.750 --> 03:55.800 You can experiment with this 03:55.800 --> 03:57.930 and see what type of results you get. 03:57.930 --> 04:01.620 Okay, I'm gonna hit Generate and hopefully this works. 04:01.620 --> 04:04.143 We'll be able to see the image in a few minutes. 04:06.180 --> 04:08.190 Okay, it's almost done. 04:08.190 --> 04:10.590 It's been going for a few minutes now. 04:10.590 --> 04:12.603 It should be done in a second. 04:14.519 --> 04:16.819 You can see here it's been about five minutes. 04:19.740 --> 04:21.930 So it is a bigger model, it does take a lot longer. 04:21.930 --> 04:24.300 But look at those results, right? 04:24.300 --> 04:27.240 Compare this, this looks really childish 04:27.240 --> 04:29.880 and not very good at all. 04:29.880 --> 04:32.490 Take a lot of prompt engineering to get this good. 04:32.490 --> 04:35.673 But look at this and this is fantastic, right? 04:36.690 --> 04:39.270 You can't argue with quality. 04:39.270 --> 04:41.280 Yeah, this is, it's definitely worth the wait. 04:41.280 --> 04:43.298 I'm glad that they finally got a model 04:43.298 --> 04:46.530 that's much better than the original. 04:46.530 --> 04:49.591 And yeah, it's not actually that difficult to change. 04:49.591 --> 04:52.530 There's still some things that aren't compatible, 04:52.530 --> 04:54.420 but there's still some bugs in the system, 04:54.420 --> 04:56.880 but it generally works incredibly well. 04:56.880 --> 04:58.113 So enjoy.