WEBVTT 00:00.935 --> 00:04.800 -: All right, let me show you how to train a product 00:04.800 --> 00:06.570 into Stable Diffusion. 00:06.570 --> 00:08.673 So this is using Dream Booth. 00:09.630 --> 00:14.020 You know, this is the Shivam Shrirao. 00:15.323 --> 00:16.950 (chuckles) I can never pronounce it. 00:16.950 --> 00:21.630 This is there, you know, specific Colab notebook, 00:21.630 --> 00:23.553 which is really, really powerful. 00:24.630 --> 00:26.580 Colab, the reason we use this 00:26.580 --> 00:29.070 is it actually gives us access to a GPU, 00:29.070 --> 00:31.470 which is needed in order to run this. 00:31.470 --> 00:33.570 And even with the GPU, it takes quite a while. 00:33.570 --> 00:36.120 So buckle up. 00:36.120 --> 00:39.120 So we're gonna go through, first thing you wanna do, 00:39.120 --> 00:40.380 I'm not actually gonna run this 00:40.380 --> 00:41.400 'cause it'll take a long time. 00:41.400 --> 00:42.870 I've already run it in the background, 00:42.870 --> 00:44.910 but I'll kind of walk you through it. 00:44.910 --> 00:48.510 So first you need to kind of install the code 00:48.510 --> 00:53.510 to train Dream Booth and the Diffusers library. 00:53.918 --> 00:57.450 You also need to log into Hugging Face. 00:57.450 --> 01:01.620 So you need to put your HuggingFace token in here. 01:01.620 --> 01:04.470 You can get that when by logging in. 01:04.470 --> 01:06.750 I'm going to, yeah, I've already changed mine here, 01:06.750 --> 01:08.670 so you know, yours would work, 01:08.670 --> 01:11.840 but if you try this, it won't work. 01:11.840 --> 01:14.850 But you'll have your own and you can just kind of 01:14.850 --> 01:19.320 visit Having Face and go to the API section to get that. 01:19.320 --> 01:21.546 I don't recommend saving the model with gdrive, 01:21.546 --> 01:23.340 it takes up four to five gigs, 01:23.340 --> 01:25.980 so it can take up quite a lot of your gdrive, 01:25.980 --> 01:28.480 but if you're gonna use it again, you can do that. 01:29.998 --> 01:32.700 The model that you're using is here. 01:32.700 --> 01:36.513 I use the Stable Diffusion version 1.5. 01:38.280 --> 01:39.900 You know, I think there are later versions, 01:39.900 --> 01:42.960 I don't know if Dream Booth specifically works them, 01:42.960 --> 01:46.080 but this is the one I found works pretty well. 01:46.080 --> 01:48.720 And then you just wanna kind of create a directory 01:48.720 --> 01:50.850 for outputting the weights. 01:50.850 --> 01:53.610 Alright, so I'm gonna start training 01:53.610 --> 01:56.190 and there's a few things you need to change 01:56.190 --> 01:57.023 before you do that. 01:57.023 --> 02:00.450 So one is like, at least specifically here, you know, 02:00.450 --> 02:02.730 if you're training on a person, 02:02.730 --> 02:05.043 like if you're making one of those AI photo, 02:06.510 --> 02:08.970 you know, photo profile things 02:08.970 --> 02:12.240 where you train yourself into the AI 02:12.240 --> 02:14.580 and then put yourself into different situations, 02:14.580 --> 02:18.690 then you would use this class instance of a person. 02:18.690 --> 02:21.510 But in this case, we're doing a product 02:21.510 --> 02:23.797 and then we're calling it the product zwx. 02:26.130 --> 02:29.640 So that's what we're going to use in the actual prompt, zwx. 02:29.640 --> 02:32.760 And you just run that and then you can upload your photos. 02:32.760 --> 02:37.760 So here I have, you know, all the photos that I've uploaded 02:38.970 --> 02:43.970 and it's actually takes a little bit of time to do that. 02:44.130 --> 02:45.461 But you know, that's done. 02:45.461 --> 02:48.480 You could also, if you know how to structure it, 02:48.480 --> 02:50.820 you can actually upload it into the instant data directory. 02:50.820 --> 02:54.540 But, you know, if you're not as technical, 02:54.540 --> 02:58.620 then just run this and click "choose files". 02:58.620 --> 03:00.960 All, right so this is where 03:00.960 --> 03:03.240 the actual training happens, right? 03:03.240 --> 03:06.750 So, you know, put the model name in, et cetera. 03:06.750 --> 03:09.090 You don't need to change anything around this. 03:09.090 --> 03:12.360 Just make sure that this is a sample prompt. 03:12.360 --> 03:15.270 So this is what is going to use to run afterwards. 03:15.270 --> 03:17.850 Although obviously you can run your own prompts 03:17.850 --> 03:20.370 and you just need to kind of make sure that fits. 03:20.370 --> 03:25.290 So I have the product as zwx or dwx and you run that 03:25.290 --> 03:26.730 and then it starts the training. 03:26.730 --> 03:29.823 So it took a little bit of a while. 03:31.020 --> 03:34.250 In this case I think it was, you know, 14... 03:35.460 --> 03:37.440 Yeah, 26 minutes. 03:37.440 --> 03:41.343 So not terrible, especially for how valuable it is. 03:43.050 --> 03:44.430 And then this is where you kind of run it. 03:44.430 --> 03:45.840 So if you already have some weights, 03:45.840 --> 03:48.120 if you've saved them in the past, you know, 03:48.120 --> 03:51.000 you could use them here. 03:51.000 --> 03:53.340 But this is where it really gets interesting. 03:53.340 --> 03:55.200 So what I trained it on 03:55.200 --> 03:57.030 were pictures of my daughter's purse. 03:57.030 --> 03:58.533 So she's three, right? 03:59.400 --> 04:03.960 And she has like this little kinda purse that she got 04:03.960 --> 04:06.660 with sequins, purple and turquoise. 04:06.660 --> 04:08.820 And the really cool thing here is that 04:08.820 --> 04:11.280 these are not real pictures of the purse, right? 04:11.280 --> 04:12.870 Like I uploaded them earlier. 04:12.870 --> 04:17.870 These are actual kind of, AI generated photos 04:17.880 --> 04:19.830 and they look similar to the ones I've uploaded 04:19.830 --> 04:23.730 because that's kind of in the training data now. 04:23.730 --> 04:27.180 Like I uploaded one that was next to my kitchen tiles, 04:27.180 --> 04:29.610 one on a wooden surface, et cetera. 04:29.610 --> 04:32.370 You wanna provide different formats 04:32.370 --> 04:36.360 and the more different scenarios and angles 04:36.360 --> 04:38.580 you put the product in, the better 04:38.580 --> 04:41.400 because then it's gonna be able to replicate that. 04:41.400 --> 04:44.070 Cool, so that kinda shows that it works, 04:44.070 --> 04:46.620 but then the way you actually use it is 04:46.620 --> 04:48.900 when you run inference down here. 04:48.900 --> 04:50.504 So you can set a seed 04:50.504 --> 04:53.400 and that makes it reproducible, so you run the same thing 04:53.400 --> 04:55.080 and it'll get the same result. 04:55.080 --> 04:59.040 But what I wanted to do, just to impress my daughter 04:59.040 --> 05:00.870 was have Princess Elsa from Frozen 05:00.870 --> 05:05.250 carrying zwx product, right? 05:05.250 --> 05:06.660 'Cause I've negative prompts, 05:06.660 --> 05:09.540 you can change the guidance scale, all these other things, 05:09.540 --> 05:12.180 but you can see that, you know, 05:12.180 --> 05:13.770 in some cases it doesn't get it right. 05:13.770 --> 05:16.683 but in other cases it kind of gets a bit creepy. 05:18.000 --> 05:20.040 Here it kind of looks like the purse, 05:20.040 --> 05:21.600 but then some of 'em turn out great. 05:21.600 --> 05:24.150 So you can see here that this really looks 05:24.150 --> 05:26.550 much like the purse and then it's pretty much got 05:26.550 --> 05:28.770 Princess Elsa here as well. 05:28.770 --> 05:31.020 So, you know, this is kind of a basic prompt 05:31.020 --> 05:32.640 and you can do more fine tuning on this. 05:32.640 --> 05:34.530 You can provide more images, 05:34.530 --> 05:36.660 you can actually do this with multiple products. 05:36.660 --> 05:39.960 But just wanted to kinda show you how you could do this. 05:39.960 --> 05:44.070 You know, if you have this use case of generating ads 05:44.070 --> 05:49.070 for a client or, you know, generating photos 05:49.290 --> 05:52.623 where you actually have to have the real product inserted, 05:53.970 --> 05:58.053 now you can train the AI on this specific concept. 06:00.121 --> 06:00.954 And that's it. 06:00.954 --> 06:04.740 So this takes up a lot of space and you know, 06:04.740 --> 06:08.610 you can just about run it on a Google Colab for free. 06:08.610 --> 06:11.640 But you know, typically once you're doing this more often 06:11.640 --> 06:15.780 you would do this on your own GPU enabled computer 06:15.780 --> 06:17.580 or in the cloud. 06:17.580 --> 06:21.240 And this is basically the same code 06:21.240 --> 06:25.890 that these AI profile photo companies use, 06:25.890 --> 06:27.660 but obviously, we're using it 06:27.660 --> 06:29.910 to train the product in this case. 06:29.910 --> 06:31.760 So, yeah, hopefully that was helpful. 06:32.640 --> 06:34.770 It's a really powerful use case 06:34.770 --> 06:37.740 and obviously if you can kind of get it to work 06:37.740 --> 06:39.690 for the more powerful Stable Diffusions 06:39.690 --> 06:41.987 or as the models get better, 06:41.987 --> 06:44.703 then you can be much more flexible with this.