WEBVTT 1 00:00:00.510 --> 00:00:03.000 Now, back in the chat window, 2 00:00:03.000 --> 00:00:04.980 which is arguably the place you'll spend 3 00:00:04.980 --> 00:00:08.400 most of your time in, and back in user mode. 4 00:00:08.400 --> 00:00:12.150 I wanna take a closer look at this plus button here 5 00:00:12.150 --> 00:00:16.770 because this button allows you to attach images 6 00:00:16.770 --> 00:00:19.080 or files to your chat. 7 00:00:19.080 --> 00:00:21.330 Now, in order to work with images, 8 00:00:21.330 --> 00:00:25.050 you need to work with a model that supports images 9 00:00:25.050 --> 00:00:27.570 that's multimodal, and you can tell 10 00:00:27.570 --> 00:00:29.940 whether that's the case for a given model 11 00:00:29.940 --> 00:00:33.213 by checking if this icon here is there. 12 00:00:34.260 --> 00:00:38.460 You can also go to your app settings and the model search. 13 00:00:38.460 --> 00:00:41.430 And there if you select a model, either one you have 14 00:00:41.430 --> 00:00:44.430 or also models you didn't download yet, 15 00:00:44.430 --> 00:00:47.220 you will also see that icon there. 16 00:00:47.220 --> 00:00:48.600 And if you hover over that icon, 17 00:00:48.600 --> 00:00:50.460 you'll also see this info text. 18 00:00:50.460 --> 00:00:53.550 This model can process image inputs. 19 00:00:53.550 --> 00:00:55.860 And you of course must be working with a model 20 00:00:55.860 --> 00:00:59.580 that does support image inputs in order to be able to upload 21 00:00:59.580 --> 00:01:01.590 and use an image. 22 00:01:01.590 --> 00:01:04.410 So for example, here, I'll open a new chat 23 00:01:04.410 --> 00:01:06.810 and I will attach an image here. 24 00:01:06.810 --> 00:01:08.730 And the image I do want to attach 25 00:01:08.730 --> 00:01:12.303 is a car wash receipt I have. 26 00:01:14.070 --> 00:01:17.700 This is the image, it's a pretty hard to read receipt 27 00:01:17.700 --> 00:01:20.190 of me washing my car or of the place 28 00:01:20.190 --> 00:01:21.450 where I did wash my car. 29 00:01:21.450 --> 00:01:23.040 It's in German, but in the end, 30 00:01:23.040 --> 00:01:26.250 it contains the amount I paid and some other information. 31 00:01:26.250 --> 00:01:28.470 And I can upload this here. 32 00:01:28.470 --> 00:01:30.060 And I could then use this model, 33 00:01:30.060 --> 00:01:32.310 since this Gemma 3 model, 34 00:01:32.310 --> 00:01:34.050 this 12 billion parameters model 35 00:01:34.050 --> 00:01:36.060 does support image processing. 36 00:01:36.060 --> 00:01:39.990 I could ask it to, "Please extract the core 37 00:01:39.990 --> 00:01:44.067 information from this uploaded image." 38 00:01:44.910 --> 00:01:47.460 And we could see whether it's able to do that or not. 39 00:01:47.460 --> 00:01:50.340 It is technically able to parse images, 40 00:01:50.340 --> 00:01:53.610 but let's see if it can read the content from this image, 41 00:01:53.610 --> 00:01:56.820 since this is really bad quality. 42 00:01:56.820 --> 00:02:00.060 Now, processing images can take a bit longer, 43 00:02:00.060 --> 00:02:03.960 but it then goes ahead and this indeed doesn't look too bad. 44 00:02:03.960 --> 00:02:06.960 Now, there are some mistakes in here, 45 00:02:06.960 --> 00:02:10.170 like the street name is not totally correct, 46 00:02:10.170 --> 00:02:13.890 but the amount I paid, for example, is correct. 47 00:02:13.890 --> 00:02:15.390 And that's of course pretty impressive 48 00:02:15.390 --> 00:02:17.880 since this is just a small model running locally 49 00:02:17.880 --> 00:02:22.770 on our system and the quality here is not great. 50 00:02:22.770 --> 00:02:27.750 By the way, for comparison, here I am using that same image 51 00:02:27.750 --> 00:02:31.470 with the more powerful 27 billion parameters 52 00:02:31.470 --> 00:02:32.730 Gemma 3 model. 53 00:02:32.730 --> 00:02:35.940 And at least here for me, it does extract 54 00:02:35.940 --> 00:02:38.250 all that information correctly. 55 00:02:38.250 --> 00:02:42.060 So using these models for OCR tasks like this, 56 00:02:42.060 --> 00:02:44.010 for extracting text from images, 57 00:02:44.010 --> 00:02:46.533 can be one very useful use case.