WEBVTT 00:00.510 --> 00:03.510 -: Hey, I'm gonna show you how to make AI profile pictures 00:03.510 --> 00:04.890 like you might've seen on Twitter. 00:04.890 --> 00:08.340 So there are a few of these tools out there that do this. 00:08.340 --> 00:10.230 And to be honest, it's probably better 00:10.230 --> 00:12.180 and easier just to use those tools, 00:12.180 --> 00:14.280 especially if you don't have access to a GPU. 00:14.280 --> 00:17.370 But you know, I think there's a lot you can do 00:17.370 --> 00:19.500 in terms of flexibility once you understand 00:19.500 --> 00:21.540 this format and how it works. 00:21.540 --> 00:22.890 We're using Dream Booth. 00:22.890 --> 00:25.020 A Dream booth is a way to fine tune 00:25.020 --> 00:27.120 the Stable Diffusion library 00:27.120 --> 00:30.300 to train it on your own images. 00:30.300 --> 00:32.130 So you could train it on a specific style 00:32.130 --> 00:34.320 or you could train it on pictures of yourself. 00:34.320 --> 00:35.790 Like I have. 00:35.790 --> 00:39.720 The main thing to note here, I'm just gonna connect, 00:39.720 --> 00:44.520 is when you connect, you wanna view your resources 00:44.520 --> 00:46.210 and change runtime type 00:47.130 --> 00:50.190 and you need to make sure that you're running with a GPU. 00:50.190 --> 00:52.470 So the reason why you need this is 00:52.470 --> 00:55.830 you're running Stable Diffusion and also training the model 00:55.830 --> 00:57.930 and that is resource intensive. 00:57.930 --> 00:59.700 So I'm not actually gonna run this, 00:59.700 --> 01:01.410 I've run it previously, 01:01.410 --> 01:03.120 just 'cause it takes quite a long time. 01:03.120 --> 01:04.890 But you'll have access to this code 01:04.890 --> 01:07.680 to be able to run it in a couple hours yourself. 01:07.680 --> 01:12.210 So we're just, you know, first thing just downloading 01:12.210 --> 01:13.950 the install requirements. 01:13.950 --> 01:15.090 So that's pretty quick. 01:15.090 --> 01:17.130 And then you need a HuggingFace token. 01:17.130 --> 01:18.513 So this one is mine. 01:20.075 --> 01:22.050 I'm gonna delete this afterwards 01:22.050 --> 01:23.730 just so you guys can't access. (chuckles) 01:23.730 --> 01:28.530 But in order to get it, you could just go into Hugging Face. 01:29.790 --> 01:33.930 This is the original Stable Diffusion, the 1.5 version. 01:33.930 --> 01:37.470 And then I'm just gonna go to the settings. 01:37.470 --> 01:40.410 I think first you have to accept this model card, 01:40.410 --> 01:42.510 which I've already done so it's not showing up for me. 01:42.510 --> 01:46.140 But if you go to Access Tokens 01:46.140 --> 01:49.680 and then you know, you can create a new access token here. 01:49.680 --> 01:52.410 So that's where you would run this. 01:52.410 --> 01:55.140 And then if you wanna invalidate and refresh, 01:55.140 --> 01:56.853 you can do that here as well. 01:58.350 --> 02:00.090 And then this token will stop working. 02:00.090 --> 02:02.100 So that's fine 'cause I'm not running this. 02:02.100 --> 02:03.960 The model name that you're using 02:03.960 --> 02:05.670 is Stable Diffusion 1.5. 02:05.670 --> 02:07.800 And then you have like an output directory, 02:07.800 --> 02:10.650 which is gonna kinda store this locally. 02:10.650 --> 02:15.240 There's just enough RAM and storage in the computer 02:15.240 --> 02:17.970 that Google CoLab gives you to store all this. 02:17.970 --> 02:20.460 But it's kind of getting to the limits. 02:20.460 --> 02:24.060 Okay, so the main thing you need to do is 02:24.060 --> 02:26.730 you need to edit this part here so you know, 02:26.730 --> 02:28.680 you can put in different concepts. 02:28.680 --> 02:32.100 I would say only do one concept at a time 02:32.100 --> 02:35.730 and then like generate the image and then cut out that part. 02:35.730 --> 02:37.980 Because when you try and combine multiple concepts, 02:37.980 --> 02:39.660 it tends not to work very well 02:39.660 --> 02:41.430 and takes a really long time for training. 02:41.430 --> 02:42.750 So here we go. 02:42.750 --> 02:44.700 What we're doing is, the instance prompt 02:44.700 --> 02:47.640 is Photo of ukj person. 02:47.640 --> 02:49.110 And this is a type of class. 02:49.110 --> 02:50.580 So if you're doing a style, 02:50.580 --> 02:51.930 like a certain type of art style, 02:51.930 --> 02:53.400 you could change that here. 02:53.400 --> 02:58.290 And then the reason why we use ukj or you know, zwx 02:58.290 --> 03:00.090 is because that's like pretty far away 03:00.090 --> 03:01.650 from an actual word, right? 03:01.650 --> 03:04.680 It's not a real word and so we won't get confused. 03:04.680 --> 03:06.780 And you know, there's different types of classes. 03:06.780 --> 03:10.350 So like this product person or you could have style. 03:10.350 --> 03:14.010 Alright, so that's just gonna run and when you run that, 03:14.010 --> 03:15.930 it will do the JSON dump here, 03:15.930 --> 03:18.300 to kind of use this concept list when it's training. 03:18.300 --> 03:22.170 And now what you can do is upload the actual images. 03:22.170 --> 03:25.530 And what I've done, I just went on my Facebook account 03:25.530 --> 03:29.070 and downloaded 30, I think it was 30. 03:29.070 --> 03:31.680 And actually only 20 images of myself, 03:31.680 --> 03:33.090 just screen shotted them. 03:33.090 --> 03:34.830 And then I uploaded them. 03:34.830 --> 03:36.450 I would say typically you need at least 10, 03:36.450 --> 03:38.430 otherwise it doesn't work very well. 03:38.430 --> 03:40.500 20 is fine, 30 is great. 03:40.500 --> 03:43.080 The more you upload, the longer it takes, obviously. 03:43.080 --> 03:44.610 But about 20 was fine. 03:44.610 --> 03:48.010 The main thing to make sure is that the images are all like 03:48.900 --> 03:51.420 very similar or not very similar, 03:51.420 --> 03:53.550 so you wanna make sure that the images 03:53.550 --> 03:55.830 have some diversity, you know, some full body, 03:55.830 --> 03:58.530 some close up, you know, some of you laughing, 03:58.530 --> 04:00.060 some of you wearing a hat, 04:00.060 --> 04:01.980 wearing different clothes, in different locations. 04:01.980 --> 04:04.860 I think that tends to improve the diversity 04:04.860 --> 04:06.810 because then it can kind of see the common threads 04:06.810 --> 04:08.880 of like what makes you, you 04:08.880 --> 04:12.120 when it kind of trains. 04:12.120 --> 04:14.370 This is the actual training code here. 04:14.370 --> 04:18.660 So nothing to change other than the sample_prompt. 04:18.660 --> 04:21.420 And so what I wanted to do is I wanted to make 04:21.420 --> 04:23.310 an illustration of myself in the style of 04:23.310 --> 04:25.800 SpiderMan: into the Spiderverse. 04:25.800 --> 04:27.390 I generated this prompt. 04:27.390 --> 04:29.430 I would say like this is really good 04:29.430 --> 04:31.620 to combine fine tuning with prompting. 04:31.620 --> 04:34.140 So this is an actual prompt that I used to use 04:34.140 --> 04:37.560 in order to make Into The Spiderverse type images. 04:37.560 --> 04:40.260 And then this is kind of like doubling up, right? 04:40.260 --> 04:41.700 So you know, we have the prompt 04:41.700 --> 04:43.650 and then we have the image of me 04:43.650 --> 04:45.630 and then, you know, that way 04:45.630 --> 04:48.630 I don't have to like kind of describe myself in this prompt. 04:48.630 --> 04:51.690 You can actually get me from the training data. 04:51.690 --> 04:53.550 Cool, so you can also save the weights 04:53.550 --> 04:54.870 and reuse them if you want. 04:54.870 --> 04:56.280 I haven't saved them, in this case, 04:56.280 --> 04:58.290 it takes up quite a lot of space 04:58.290 --> 05:01.200 in your Google drive or locally. 05:01.200 --> 05:03.330 And then when you run this, 05:03.330 --> 05:06.930 then it's generating the example images after the training. 05:06.930 --> 05:10.770 So just to see here, this took about 10, 05:10.770 --> 05:13.140 I guess this is 10 minutes. 05:13.140 --> 05:15.325 It's not, oh no, sorry, it's 10 hours. Yeah. 05:15.325 --> 05:17.520 (chuckles) So it's quite a while I was running it. 05:17.520 --> 05:20.430 Yeah, I think it's 10 minutes, 05:20.430 --> 05:23.313 but sometimes it takes a long time to run. 05:24.870 --> 05:27.540 I've run it before and it took half a day. 05:27.540 --> 05:29.550 And so I think that's probably the longest. 05:29.550 --> 05:31.110 You just make sure you have the GPU 05:31.110 --> 05:32.880 and that's what makes it faster. 05:32.880 --> 05:34.590 Cool. So it did a pretty good job, right? 05:34.590 --> 05:36.570 Like this is a decent image of me, 05:36.570 --> 05:39.630 obviously this is kind of cheating, showing me in a mask, 05:39.630 --> 05:42.960 but you know, this is kind of, taking my beard 05:42.960 --> 05:45.420 and my eyebrows and my hair and stuff 05:45.420 --> 05:48.960 and yeah, I think it's done a good job of matching the style 05:48.960 --> 05:51.990 and putting me in into the Spiderverse. 05:51.990 --> 05:53.640 You can also save the weights and kind of 05:53.640 --> 05:55.560 use that for... 05:55.560 --> 05:59.340 At some of the web UIs, like automatic, if you want. 05:59.340 --> 06:03.270 But you can also just keep running it in this Google CoLab. 06:03.270 --> 06:06.720 So at this part here is to actually kind of run the pipeline 06:06.720 --> 06:09.390 and then you can set a seed from here. 06:09.390 --> 06:11.370 And that's just so you get the same image every time 06:11.370 --> 06:12.660 if you use the same prompt. 06:12.660 --> 06:15.780 And then I've used at this specific prompt here, 06:15.780 --> 06:19.290 illustration ukj person in the style of SpiderMan, 06:19.290 --> 06:20.640 blah, blah, blah, right? 06:20.640 --> 06:22.560 So that's what it's come back with. 06:22.560 --> 06:25.350 And you can edit this so you can say, you know, 06:25.350 --> 06:26.910 I want guidance scale. 06:26.910 --> 06:30.450 Like you know, the higher the guidance scale, 06:30.450 --> 06:31.800 the closer it is to the prompt, 06:31.800 --> 06:35.100 the lower, the more kind of creative it can be. 06:35.100 --> 06:37.230 There's also number of inference steps. 06:37.230 --> 06:39.420 So, you know, if these steps are higher, 06:39.420 --> 06:43.140 it tends to be higher quality images, but not always. 06:43.140 --> 06:45.690 I think it's good to experiment with this. 06:45.690 --> 06:47.070 And then number of samples as well. 06:47.070 --> 06:49.560 So I've generated four images here 06:49.560 --> 06:52.413 and yeah, you can see it's done pretty cool job here. 06:53.640 --> 06:56.163 Cool. So, that's how this works.