WEBVTT 00:00:00.000 --> 00:00:32.230 Here's the easy way to create AI videos. In this tutorial, I'm showing you exactly how I generated this realistic forty four second long fight scene and how I made the seventy one second long short film, and I'll show you how to extend them even more to be as long as you want. This is by far the simplest and easiest method to generate long AI videos. The first step is to use AI to generate a storyboard just like this one. It's gonna divide our long AI videos into these separate scenes. The tool we're gonna use for this is the newest GPT two image model. 00:00:32.390 --> 00:00:40.710 What this tool is amazing at is reasoning and generating text, and that's gonna come in really handy for us when we're trying to generate these lawn storyboards. 00:00:40.905 --> 00:00:44.665 To use GPT image two, I'm inside Higgs Field AI, 00:00:44.825 --> 00:00:50.185 and you can find it on the home page, or we can go to the list of image generators 00:00:50.265 --> 00:00:58.870 and find the new GPT image two model. For the first example I'm creating, I'm gonna generate an lawn AI video based around these two photos 00:00:59.030 --> 00:01:05.030 of the scientist in the hazmat suit and his robot companion who are exploring this toxic forest. 00:01:05.110 --> 00:01:06.710 So we have the GPT 00:01:06.710 --> 00:01:12.945 image two model chosen, and what I'm gonna do is upload those two reference images of my characters. 00:01:13.265 --> 00:01:25.720 Let's put those inside here. Next, I'm gonna ask the g p t image two model to create a full storyboard for us. The prompt can be super, super simple. Just a one sentence description is gonna be enough to create a full storyboard. 00:01:25.800 --> 00:01:30.520 Now you have some other options, uh, like choosing the quality, 00:01:30.920 --> 00:01:33.000 the resolution of the storyboard, 00:01:33.735 --> 00:01:41.095 and also the aspect ratio. I'd recommend setting this to 16 to nine. And then we just have to generate the AI storyboard. 00:01:41.175 --> 00:01:49.560 This is the full built out story. It's got 12 complete shots in it that tell a full story, and underneath each individual shot, 00:01:49.720 --> 00:01:59.080 there's a small text description. Those little text descriptions are gonna be super useful later on because we're gonna use those as part of the prompts for our AI video generator. 00:01:59.475 --> 00:02:05.155 I did notice sometimes that when you're generating a bunch of separate panels like this, there is some repetition. 00:02:05.235 --> 00:02:18.310 So if you look at this image in panel three, which is the robot gesturing upwards, it ends up being the same image as shown inside panel 11. So we need to make a minor edit to the storyboard so we don't have this repetition. 00:02:18.390 --> 00:02:21.750 So inside Hicksfield, what I'm gonna do is hit reference, 00:02:21.990 --> 00:02:24.790 which is gonna let me edit this storyboard. 00:02:25.510 --> 00:02:32.715 And so there's that storyboard added on into our image references. Let's delete all the other ones, um, and also 00:02:33.035 --> 00:02:36.795 change the prompt. What we wanna do is change 00:02:36.955 --> 00:02:54.750 panel 11 so that it's not a repetition of panel three, and that's basically exactly what I'm gonna tell the AI to do. Just adjust shot 11 so it's not a repeat of shot number three. And looking at the result, it's swapped shot 11 to one of the scientists, uh, inspecting some kind of a toolkit. 00:02:55.725 --> 00:03:08.190 So now that we have the storyboard sequence, how do we leverage this and turn it into a lawn AI video? We're gonna use a AI video model called c dance two point o. So if we have a storyboard, 00:03:08.270 --> 00:03:13.950 and what cdance can do for us is to animate all of these in a single video generation 00:03:14.190 --> 00:03:28.045 at the same time. To use the cdance video model, I'm gonna look through the video models on Higgs field, and the top one is cdance two point o. So for this AI video model, the maximum duration of each video generation 00:03:28.205 --> 00:03:30.365 happens to be fifteen seconds. 00:03:30.525 --> 00:03:33.965 So then how do we create a super long AI video sequence 00:03:34.045 --> 00:03:36.045 of our entire storyboard? 00:03:36.290 --> 00:03:39.090 Well, if we try to fit this entire storyboard, 00:03:39.090 --> 00:03:43.650 all 12 shots into a single fifteen second video, 00:03:43.730 --> 00:03:51.885 there just isn't enough time inside those fifteen seconds to animate this entire storyboard. So I'm gonna split the storyboard up 00:03:52.125 --> 00:03:54.845 and crop out each individual 00:03:54.845 --> 00:04:01.885 role. So this is the first role I've cropped out where I'm just taking the first four shots of that storyboard. 00:04:01.885 --> 00:04:27.445 And what we're gonna do is animate these four shots inside the fifteen second video sequence, and it should be able to produce a pretty good result for us. A quick technical point, you will need to layer the cropped roll on top of a 16 by nine image. This is just to make sure that Higgs Field can actually use this image as a reference. For the prompt that I'm gonna use, what I'm gonna rely on is the pregenerated 00:04:27.525 --> 00:04:37.160 text descriptions that's already inside the storyboard. So what we're gonna do is tell it to generate a scene using the shots in the uploaded film storyboard. 00:04:37.240 --> 00:04:41.240 Then for each of the shots inside this fifteen second 00:04:41.480 --> 00:04:56.025 video generation, I'm gonna give it a time frame of when I want the shot to happen. For example, the first four seconds. And then for the description of what happens in the shot, I'm basically just gonna copy the text description that's already described 00:04:56.905 --> 00:04:58.425 in the storyboard, 00:04:58.425 --> 00:05:07.880 and I'm gonna fill out the rest of the prompt using the same method. You'll also wanna add in this line at the bottom that says no music and no subtitles. 00:05:07.880 --> 00:05:20.745 I found that that just makes it much easier later on when we're putting everything together. Before we go and use this prompt though, there's an really important addition that we need to add into this to preserve character consistency. 00:05:21.065 --> 00:05:42.310 So if we just used the prompt exactly written like this, we'll get results like this, which does basically animate the scene as expected except the robot doesn't look quite the same. For starters in this sequence, the robot's legs look super long for some reason. He should be much shorter than that, and there's a lot of different variation 00:05:42.390 --> 00:06:01.830 in the way that the character looks between different video generations. We need an additional character reference sheet so that when we generate the long AI video sequence of them, they actually stay consistent throughout the entire scene. To create a character reference sheet, I'm still gonna use the GPT image two model, and I'm gonna upload the original 00:06:01.830 --> 00:06:29.360 shot of my robot in the forest. And then in the prompt, I'm gonna tell it to create a character reference sheet. I'm gonna put the prompt in the description so you can go and copy it, and the prompt is gonna give us this really useful character reference sheet to help us maintain consistency inside the videos. Now that we have all these different assets, we can start putting together our long AI video sequence. So first off, I'm gonna upload into Higgs Field the first four shots of our storyboard 00:06:29.360 --> 00:06:32.480 along with a character reference sheet of our robot. 00:06:32.640 --> 00:06:39.040 Let's drag that inside there. And inside the prompt, we're gonna write it exactly like I previously described. 00:06:39.605 --> 00:06:45.285 So first telling it to generate a scene using the shots inside the uploaded film storyboard. 00:06:45.445 --> 00:06:48.725 And here, I'm actually gonna make a reference 00:06:48.885 --> 00:06:51.125 to the film storyboard, 00:06:51.285 --> 00:06:57.640 uh, uploaded image. So I'll type at, which lets me tag different references, 00:06:57.720 --> 00:07:02.920 and then I'm gonna write out each of the individual scenes. Just basically copy in the 00:07:03.240 --> 00:07:05.240 text descriptions already in the storyboard. 00:07:05.765 --> 00:07:09.045 And then inside this prompt, I'm also gonna need to reference 00:07:09.125 --> 00:07:10.645 our character sheet 00:07:10.805 --> 00:07:18.325 for the robot as well so that the AI video knows exactly what the robot should look like. So here, I'll add an additional tag 00:07:18.990 --> 00:07:21.470 next to the word robot 00:07:21.470 --> 00:07:28.110 for our character sheet. I'm actually also gonna do this in a few other places as well. 00:07:28.590 --> 00:07:34.030 And then in the settings, I'm just gonna make sure that I'm using the full fifteen second 00:07:33.155 --> 00:07:34.515 video duration. 00:07:35.395 --> 00:07:38.355 Now let's generate the video and see what it looks like. 00:07:53.550 --> 00:07:56.270 It actually looks like it animated five 00:07:56.430 --> 00:08:07.075 shots here instead of the four shots that are prompted for, but that's not a bad problem to have. It still followed all the shots inside the storyboard accurately 00:08:07.075 --> 00:08:22.400 and the robot looks great. And then using the same exact method, I went ahead and animated the rest of the storyboard. So here's the animation sequence for roll number two of the storyboard, which is shots five through eight. 00:08:24.240 --> 00:08:26.800 And here are the last four shots of the storyboard. 00:08:30.575 --> 00:08:51.340 And now we have three separate fifteen second video clips that are in sequence to each other, which means that we can combine them together into a full forty five second AI video sequence, but it doesn't stop here. We can actually extend our storyboard as many times as we want to. So looking at the sequence of shots inside the storyboard, 00:08:51.340 --> 00:09:11.570 what if we wanted to extend this so that they start exploring the forest even more? What we can do is actually use GPT image two to generate the next 12 panels of the storyboard as well. So here's what we're gonna do. Inside the g p t two image generator, we're gonna upload the original storyboard reference. 00:09:11.810 --> 00:09:15.490 So this is the first 12 shots that we just animated. 00:09:15.490 --> 00:09:23.385 And then I'm also gonna upload a character reference sheet for my scientist and also the character reference sheet that I generated for the robot. 00:09:23.545 --> 00:09:32.745 Then inside the prompt, I'm gonna ask it to simply generate the next 12 panels of the storyboard, and here's the prompt I'm gonna use. Generate the next page of the storyboard 00:09:32.960 --> 00:09:38.320 from image one, which continues the story with 12 panels using the uploaded images. 00:09:38.400 --> 00:09:41.360 The robot reveals hidden knowledge of the forest, 00:09:41.600 --> 00:09:46.880 guiding the scientists to a deeper, more dangerous core where the source of the toxic outbreak lies. 00:09:47.555 --> 00:09:50.995 And this is the extended storyboard that GPT 00:09:50.995 --> 00:10:03.480 image two has created for us. And then I can use the same exact technique of animating four separate shots at the same time to turn this storyboard into three fifteen second 00:10:03.480 --> 00:10:04.200 video 00:10:04.440 --> 00:10:06.040 generations as well. 00:10:08.040 --> 00:10:11.400 And if you do the math, using the two storyboard pages, 00:10:11.655 --> 00:10:17.735 we'll end up with six fifteen second video clips, which is a ninety second video sequence. 00:10:18.215 --> 00:10:22.695 I found that some of the sequences that I generated were a little bit repetitive, 00:10:23.175 --> 00:10:26.375 so I ended up trimming it down to a seventy one second 00:10:26.760 --> 00:10:28.040 video instead. 00:10:28.440 --> 00:10:33.960 But you can use this method and extend your videos for as long as you want. 00:10:35.240 --> 00:10:41.720 I'm gonna put a link in the description for Higgs Field AI if you wanna go and generate your own long video sequences 00:10:41.455 --> 00:10:57.940 using GPT image two and c dense two point o. Now, one of the challenges we're gonna run into this especially when animating a more dynamic action packed scenes, for example, this fight scene that I created, is that because of how much action there is, I found that if we animate 00:10:58.180 --> 00:11:00.660 each of the roles inside the storyboard 00:11:00.740 --> 00:11:02.820 separately by themselves, 00:11:03.140 --> 00:11:06.820 it's much harder to combine them together seamlessly. 00:11:06.980 --> 00:11:09.380 So let's see how we can fix this problem. 00:11:09.945 --> 00:11:17.305 First off, for my fight scene, these are the characters and the environment that I want them to be in, and this is the storyboard 00:11:17.385 --> 00:11:24.720 and the prompt that I used inside GPT image two to create this. Now when I go and animate each row separately of the storyboard, 00:11:24.960 --> 00:11:33.040 each of the individual animations look really, really good. So this is the first role animated, and keep an eye on what happens in the last scene. 00:11:36.495 --> 00:11:41.855 The bounty hunter suddenly ambushes the female character and has her in this chokehold. 00:11:42.495 --> 00:11:49.695 Now when I go and animate the second row of the storyboard, it starts off with the first frame of them already engaged 00:11:49.695 --> 00:11:53.110 inside a fight scene. If we look at this individually, 00:11:53.110 --> 00:11:57.510 it's a really amazing looking AI video. However, if we try to combine 00:11:57.750 --> 00:12:01.270 those two separate fifteen second clips together, 00:12:01.590 --> 00:12:06.415 the transition is gonna look a little weird. So starting with the first video sequence 00:12:07.375 --> 00:12:16.975 and suddenly jumping into the second video sequence. That transition right there, it doesn't really make sense. She's in a choke hold and suddenly is free and engaged in a fight scene. 00:12:17.690 --> 00:12:27.130 So then how do we get the transitions to be seamless inside our lawn AI video? What we're gonna need to do is give the AI some extra information 00:12:27.370 --> 00:12:34.045 when generating each individual video clip. So first, I'm gonna use this tool called video frame extractor, 00:12:34.045 --> 00:12:41.165 and I'm gonna save that last image frame of our video sequence. Now when we go and generate the next four shots 00:12:41.245 --> 00:12:45.485 of our storyboard sequence, actually tell the AI to use the screenshot 00:12:45.690 --> 00:12:54.570 I just saved as the first frame and generate those next four shots starting from that initial screenshot. 00:12:54.810 --> 00:13:07.625 This is what it's gonna look like inside Higgs field first with all of our image references uploaded. And then inside the prompt, I'm gonna tell it to generate a scene using the uploaded film storyboard fight sequence 00:13:07.945 --> 00:13:10.345 starting with this image frame 00:13:10.425 --> 00:13:17.700 of the female character getting attacked. And then the rest of the prompt is written just like how it was done before, 00:13:17.860 --> 00:13:18.580 basically 00:13:18.900 --> 00:13:22.580 describing what happens inside each storyboard sequence. 00:13:22.740 --> 00:13:40.400 Now when we go and generate this scene, it should generate the next four shots starting with the initial image frames. And using this method, you can generate endless continuous shots for your AI films. Here's a quick note about using Higgs field. When you upload image references, 00:13:40.800 --> 00:13:44.880 what Higgs field is gonna do is check the eligibility 00:13:44.880 --> 00:13:46.240 of each image. 00:13:46.400 --> 00:13:49.680 This is to avoid any copyright issues. 00:13:50.000 --> 00:13:54.935 Now if you're using, like, a celebrity or a scene from a movie or something, 00:13:55.255 --> 00:13:56.615 it's gonna get denied. 00:13:57.015 --> 00:14:01.015 But sometimes when you upload images of your own characters 00:14:01.015 --> 00:14:03.255 as well, it can also get denied. 00:14:03.415 --> 00:14:13.220 So the first time that I tried to upload my reference sheet for my bounty hunter character, it actually got determined as not eligible. 00:14:13.700 --> 00:14:18.180 But then when I tried again, the next time it was determined 00:14:18.180 --> 00:14:25.705 as eligible to be used. So if you try to do this and you upload an image of yourself or your characters and it gets denied, 00:14:25.865 --> 00:14:37.590 just try uploading that image reference a few different times and eventually it might work. Let's take a look at what the full animated fight scene looks like so you can get an idea of what this method is capable of. 00:15:21.535 --> 00:15:28.095 If you also want a complete breakdown of 10 practical tips to generate the most realistic 00:15:28.095 --> 00:15:31.650 possible AI videos, go watch his guide right here.