WEBVTT 00:00.080 --> 00:03.600 All right, so we're doing fine tuning with flux. 00:03.640 --> 00:07.840 Laura, this is using the file AI API. 00:07.840 --> 00:10.720 And you can see I made a picture of myself. 00:11.080 --> 00:12.360 How did the AI know about me? 00:12.360 --> 00:16.560 I fine tuned it based on some images that I uploaded. 00:16.960 --> 00:20.040 So this one I'm not going to run because it takes a little bit of time. 00:20.040 --> 00:22.600 I'm just going to walk you through the code. 00:22.600 --> 00:28.760 But the general idea of what we're trying to go for is that you upload a zip file of images, you put 00:28.760 --> 00:36.120 in a prompt, and you prompt this new model that you fine tuned with a specific trigger word. 00:36.120 --> 00:38.760 In this case, I just chose Random letters UK. 00:39.320 --> 00:45.600 And now it knows that thanks to the fine tuning we did earlier, that UK means me right? 00:45.640 --> 00:47.840 And I uploaded a bunch of images of myself. 00:48.080 --> 00:49.640 You can run this yourself. 00:49.640 --> 00:53.400 It does cost a bit of money if you are going to run it on file. 00:53.520 --> 00:56.800 I also recommend you check out other providers as well. 00:56.800 --> 01:02.050 Replicate I think has done more for fine tuning, although it is possible to do it on four. 01:02.530 --> 01:04.210 So let me walk you through this. 01:04.770 --> 01:09.930 First you have to install the file client and load the environment variables. 01:10.090 --> 01:14.370 You can get the API key from the file dashboard. 01:14.370 --> 01:16.090 And you just need this file key. 01:16.530 --> 01:20.010 And the other thing you need is to prepare a bunch of images. 01:20.610 --> 01:22.930 Here you can upload a zip file. 01:22.930 --> 01:24.410 And I just had that locally. 01:24.410 --> 01:26.810 I just got a bunch of pictures of myself. 01:26.810 --> 01:29.610 I think it was like ten different pictures from my Facebook account. 01:29.770 --> 01:35.850 And then I downloaded those images and, uh, created a zip file on my Mac. 01:35.930 --> 01:40.490 And then I put that into this project and I called it MT Face my face. 01:40.890 --> 01:45.650 If you want to see how to do this, we can go to this URL here. 01:46.210 --> 01:50.610 So this is the Flux Pro trainer that we're using. 01:50.970 --> 01:52.690 You can see how much it costs. 01:52.690 --> 01:57.490 So for example it's about I think it's like between 2 and 4 to 2 to $6. 01:57.770 --> 01:59.170 And you can actually run it in here. 01:59.170 --> 02:02.730 So you don't necessarily need to use the API. 02:02.770 --> 02:07.930 You can just upload your images or pick a zip file and then choose your trigger word as well. 02:07.930 --> 02:11.970 If you want all the code that's up to date, like you can get the code here. 02:11.970 --> 02:13.090 So that's what I've done. 02:13.890 --> 02:17.770 And you can see this is the job that ran for me and it completed. 02:17.810 --> 02:19.970 I had a failed job before as well. 02:20.250 --> 02:27.730 Once you have this you can also just click to run inference and it's going to take you straight to the 02:27.930 --> 02:30.010 model that lets you run it fine tuned. 02:30.010 --> 02:32.210 So this is a fine tuned Flux Pro. 02:32.570 --> 02:37.130 And you just have to give the fine tuned ID which is the output you get, and you choose the fine tuning 02:37.130 --> 02:37.770 strength. 02:37.770 --> 02:40.410 So those are the two things we're going to be doing today. 02:41.210 --> 02:45.970 We just want to make a little message to keep us updated when we're running the trainer. 02:46.690 --> 02:51.610 And we're going to subscribe to this trainer, we just need to pass in a few different things. 02:51.610 --> 02:53.530 One is just a name. 02:53.610 --> 02:58.370 So you can find this fine tune again, just to test face, you want to choose the mode. 02:58.410 --> 03:00.820 I did character because I was doing a picture of me. 03:00.820 --> 03:06.180 But you could also do this for products or a specific style, or just a general fine tune, and then 03:06.180 --> 03:07.500 you choose the trigger word. 03:07.540 --> 03:09.620 A lot of people use Tok tok. 03:09.860 --> 03:11.220 I use UK. 03:11.700 --> 03:15.700 It doesn't really matter what you use, as long as it's something that won't be in the training data. 03:15.700 --> 03:21.220 So it's something that there's no real image for in the training of the client. 03:21.700 --> 03:23.580 And when you run that, you get a fine tune ID. 03:23.620 --> 03:28.740 It gives you an ID that you can follow up on and then you can submit the job. 03:28.940 --> 03:30.060 This is the other way to do it. 03:30.060 --> 03:31.220 You could submit the job. 03:31.580 --> 03:34.340 You can take this request ID that you get. 03:34.340 --> 03:36.980 And then you can get the status whenever you want. 03:37.020 --> 03:39.220 Like in this case this is completed. 03:39.260 --> 03:42.980 And it took 759 seconds here. 03:43.300 --> 03:46.460 There's also you can get the result the fine tuned ID. 03:46.460 --> 03:50.780 So when you've run the full thing and it's completed that's how you get the result. 03:51.180 --> 03:55.820 If you're doing this programmatically, then you just want to run in a loop and just check each time 03:55.820 --> 03:57.220 as the status completed yet. 03:57.220 --> 04:01.870 And once it is, then you've got your fine tuned ID that you can call back, but yeah, once you've 04:01.870 --> 04:06.790 got the fine tuned ID, you can just use that model as much as you want forever for as long as you want, 04:06.830 --> 04:11.510 and it's just something you have to pass in here to this fine tuned result here. 04:11.510 --> 04:12.510 We can actually run this. 04:12.510 --> 04:22.630 I'm not going to do another fine tuned job, but let me just run that and that and then just gonna upload 04:22.630 --> 04:29.910 the file clean here, the fine tune ID I'm just going to paste in here. 04:30.750 --> 04:31.790 That was what it was. 04:32.750 --> 04:34.150 And then I can run this. 04:34.190 --> 04:36.390 So I asked for the LinkedIn profile picture. 04:37.390 --> 04:40.830 I just need to bring in my Q update which is here. 04:41.750 --> 04:44.670 This is going to give me messages while it's running. 04:46.870 --> 04:53.670 See, when it comes back with oh, I didn't define the request library, just encrypt that. 04:54.070 --> 04:54.390 All right. 04:54.390 --> 04:57.590 So I've got all the right libraries here. 04:57.630 --> 05:00.630 Now there's another image of me. 05:01.510 --> 05:02.830 Look a bit like JD Vance. 05:03.270 --> 05:07.350 You could just say editor instead of LinkedIn. 05:07.390 --> 05:08.030 Profile picture. 05:08.030 --> 05:13.710 You say picture of you in Times Square because it knows who your CK is. 05:13.710 --> 05:20.430 Now, with this fine tune ID, then it will be able to kind of follow that prompt. 05:21.430 --> 05:21.990 There we go. 05:23.550 --> 05:28.470 Funnily enough, one of the things you find is that you tend to it tends to make all the other people 05:28.470 --> 05:30.990 in the image similar to you as well. 05:31.390 --> 05:34.870 So you can see here it's made a few very similar to me. 05:35.150 --> 05:36.750 But yeah, that's that's how it works. 05:36.750 --> 05:37.790 That's fine tuning. 05:37.830 --> 05:38.750 It's pretty straightforward. 05:38.750 --> 05:40.390 You don't have to know anything about machine learning. 05:40.390 --> 05:46.110 Thankfully, I massively prefer to do this rather than trying to figure out my own computer, although 05:46.110 --> 05:51.030 you can actually run this on your own, Mac if you have an M1 or above, or in your PC if you have a 05:51.030 --> 05:51.910 gaming PC. 05:52.110 --> 05:57.190 So these tools are open source, but I prefer to use an API wherever I can, even if it costs a couple 05:57.230 --> 05:57.790 of bucks. 05:58.590 --> 05:59.110 All right. 05:59.150 --> 06:00.510 Yeah, hopefully that's useful for you.