How I Make Realistic AI Videos with Seedance 2.5 - Full Guide Transcript, AI Summary & Key Points

00:01 Yo, what's good? It's your boy Dan. And I'm going to show you how to make realistic >> [screaming] >> This guy is insane. Here is where we like to chill. LOOK WHAT I HAVE. LET'S TEST THAT YEAH! >> [screaming] >> YEAH! >> [screaming] [music] [music] >> MANY PEOPLE ARE STILL STRUGGLING TRYING TO generate realistic looking AI footage. And that's why over the last few weeks I've been putting together a workflow of how you can achieve this.

01:30 We'll be using Higgs Field for this. In my opinion this is one of the best AI video platforms out there and they were also kind enough to sponsor this video. Okay, let's jump in. The question that I get the most is which model should I use? And this goes for both image and video. And my answer to that really is it depends. So, let me show you why it depends on what you're trying to generate.

01:51 Because if you go over to Higgs Field and you click on images, then you can see we have all of these different models. Like there's so many different models you can choose from. And it might be even a bit of an overload. So, you might automatically assume, "Oh, this is new. I should be using that model." In my opinion, you're wrong. So, let's take this example.

02:08 I have this prompt of a gangster-like chimp, and he has like diamond grills and all of that. Let's just hit generate. I've put the same prompt into six different AI image models to show you the difference between them. And you can choose for yourself which one you like the most. Because my answer to this is there are multiple good models that you can use.

02:27 You shouldn't stick to just using one. For example, I really like GPT image 2.5 flare here, Flux 2.0 Pro, even though like Flux is a really old model, and for comparison I put Flux 3 in there, too. Flux 2 is much better in case in this realism. Nano Banana 2.1, the newest model, also does a great job. And SeaDream 5.0 also looks quite good. XField Soul sometimes surprises me, but that's why you don't want to stick to just using one model and call it a day.

02:56 Images are cheap. Just generate many different models all together at the same time. Here we have another example. Again, for all of them I used the same prompt. And here you can see which one you like the most yourself. Flux 3 feels very plastic again. I really like the look of GPT image 2.5 flare, and I think Soul 2 here gives like an interesting, like more realistic look to it as if it was actually taken on like an iPhone at like the morning sunrise.

03:19 Next example, again, we're using the same prompt. In my opinion, GPT 2.5 flare looks best, but I also like the look of SeaDream 5.0 Pro here, and I think the other ones are not too bad, either. Here's example number four, and here we are really going for that disposable camera type of look. I also asked it to include this timestamp, so we have that in the image, and Soul 2 and Flux 2.0 Pro didn't do that because they're bad at generating text on your images.

03:45 But the look of Flux 2 Pro is really good. I also like GPT image 2.5 flare here, but it feels a bit too blurry if you were to zoom in. And then the last example here, I like Flux 2.0 Pro the most. The only thing it didn't do is follow up my prompt. I told the AI to give it the finger and it just put up a finger instead of putting up the middle finger, but it looks insanely good.

04:09 Like it looks like it was actually taken from a camera. So, what I'm trying to say with this is that these models are trained on like different data sets. Some of them are better at certain things. For example, for realism, Flux 2.0 Pro is still quite good. Nano Banana Pro and Nano Banana 2.1 also are still quite good. I maybe even prefer the Nano Banana Pro look over Nano Banana 2.1, but GPT image 2.5 flare surprised me over using a tool like GPT 2.5 sunburst.

04:38 And that's why throughout this video, I don't want you to stick to just using one model. Try out different generations. Image generation is cheap. And especially these older models like Flux, like Soul, like Sea Dream, they are super cheap. GPT 2.5 and Nano Banana may be a little bit more expensive, but still compared to videos, it is cheap. So, don't limit yourself to just sticking to one image model and think that that is the reason your image looks good or bad.

05:04 Let's talk about AI video models because here it makes a bit more of a difference which model you're choosing for realism. I'll be showing you four different generations and I will tell you why. We have Sea Dream 2.5 1080p. We have Minimax HD Max. We have Flux Tree and we have Sea Dream 2.5 480p draft mode. We're using draft mode because it's cheap and sometimes even for realistic footage, the draft mode, because it's lower quality, helps with that realistic aesthetic.

05:31 But you could also be using something like Kling 4.0, which I don't have access to as of recording this, but test it out. Don't limit yourself to these models. So, for the first generation, I don't use any references because that's the test that I want to do. And then the second test will include references. For all of the prompts that I'm using, including all of the files, all of the assets, everything I share in this video, I will leave a link to my school community where you can find everything per video, so you

05:58 don't have to take notes. You can just go to it. You can copy my prompts. You can analyze my style, and you can learn from that. So, the goal for this prompt is that we are doing a raw found footage feel compilation of a gang of street monkeys causing chaos in everyday human places. It's kind of like Jackass style. That is the video that we're going to do, 10 seconds long.

06:17 And then I've described like all of the different things that I wanted to do. Again, this is not about the prompt, but I want to show you what we're generating. So, here's the result for Cidence 2.5 1080p. And honestly, this is very good. Like, you might already know Cidence 2.5 is one of the best AI video models. If you want that realistic look, I would use this model if you can afford it because it's quite expensive.

06:48 Then, let's compare it to Minimax H3 Max. Unfortunately, Minimax H3 is way worse compared to Cidence when it comes to generating realistic-looking footage with no references in it. Then I have Flux 3. >> [laughter] >> Now, this one is quite good because this one will get you like 80 or 90% of the result of Cidence 2.5, but it is way cheaper. So, for these type of videos, Flux 3 works quite well.

07:31 Then we have Cidence 2.5 draft mode. This is in 480p. >> Again, this is very similar to the 1080p version, but the difference here is if you're using this draft mode, you can like regenerate over and over again while using like the cheapest type of generation, which is the 480p version. And then once you really like the result that you have, then you can upscale this in 1080p.

08:02 And then you only then pay like a one-time fee to upscale your video, which is used often to save yourself some credits. Okay, so to summarize, for image models, you just want to play around, test out multiple different image models. For video models, seed and stupid 5 remains on top. There's nothing that's going to beat it, but you might want to try out flux tree and also minimax H3 if you're after that realistic look.

08:24 And also try out clink 4, but from my first look at it, I found it less realistic than seed and stupid 5. So, not too sure about it, but I haven't tried it out myself yet. Okay, now let's move over to the next segment, which is learning how to prompt for realistic-looking images. For this, we'll be using the following method. It's called the CTLT method.

08:43 If you have no idea what it means, let me explain. The C stands for camera, the T stands for tone, the L stands for lighting, and T stands for texture. I've created this template for you, which again, you can find in all of the assets that are linked in this video. So, why are we using this formula? Basically, it is to teach yourself to think about all of these different parts of what goes into your prompt.

09:06 So, for example, you want to give the camera settings such as the lens that's been used, the framing. You want to set the mood, the atmosphere, and the emotional energy first. Then you want to describe who's in the image, what they're doing, and where it is happening. Then you can also describe the lighting and the texture, where you specify physical details that interact with the environment.

09:24 Okay, let's do this together as an example. So, for the camera, I want to have a candid 35-mm film photograph with a 12-mm fish-eye lens. You could ask your LLM to help you with this. And no worries, at the end of this template, I have this LLM which I we will go over in a second, which will help you write this prompt. But we're going to do a candid 35 mm film photograph with a 12 mm fish eye lens.

09:51 It's slightly elevated side view with a wide full body framing, tilted horizon, and a slight motion blur. Then for the tone, I want something weird. So I'm going for a frantic, absurd, and darkly comic, like an unexpected chase caught by a passerby. So basically what I'm describing is, we weren't meant to see this. It's just like someone drove by or someone passed by and took a picture of it.

10:11 Then next up, we're kind of like describing the action of what we're seeing. So I want to do a skeleton that is wearing a dress made of raw meat that is sprinting across the streets of Los Angeles at night with three dogs chasing close behind him. For the lighting, I want to do a harsh direct on camera flashlight on the skeleton. And then for the texture or the details, we're adding in matte ivory bones, glossy marbled meat with visible fat and uneven edges flapping.

10:35 Like this kind of get disgusting to describe this, but it's just describing what we want to see in terms of the details. Okay, so now if we combine this prompt, like all of these different segments, and we copy that, we go over to image, and then we create an image, and then we drop in that prompt right here. Then we can select our model. Let's do something funky.

10:54 I think Flux 2 might be good for this. I'm going to use Flux 2 Pro, and now we can select our dimensions. Let's do 16 by 9, and let's do 2K quality. Then we hit generate, and that gave us this result. Like this looks quite cool. This is what I wanted, and this is how we prompt simplified. Now I want you to know that you don't have to be a filmmaker or something like that to get prompts like this.

11:19 You could use an LLM, and I go over to my ChatGPT or my Claude. So I've put together this prompt for you. I paste it in, and then I scroll down, and here I paste in my idea. Or you just whisper flow what your idea is. So, here in this case, I could say like, "It's a bystander taking a photograph with a hard flash of a skeleton that is covered in a meat dress being chased by three dogs, and it's kind of like a image that I'm not supposed to see."

11:47 That is my idea. You hit send, and then I will help you flush out that idea using that structure. There are certain words and sentences that you can use to create unique styles. For example, this is a close skate camera type of image. If I change up the wording a little bit, I can also make some kind of like intimate documentary. I can make a motion blurred photo dump.

12:10 I can do a low curb side shot. We can do disposable camera look or a concealed observation. As you can see, these all have a totally different visual identity and style to them. The important thing is is that you just have to understand how to come up with these kind of different styles. That's why I put together this document that is seven ways to capture an idea, where we go into different structures that you can look at.

12:35 And then you can analyze these prompts to create shots like these. For example, this phone snapshot is also quite cool. So, I filled in the details for you, so you can just look at it, and you can copy these structures for your own images. Let me just give you one more demo of how I would do this with my LLM. So, we start up by copying this prompt again.

12:54 Then we go into ChatGPT, and then we add in the reference image that we want to use. For example, I have this reference image of Bruiser the robot that I want to use for my video. So, for my idea, I'm going to whisper flow Bruiser sitting outside a laundromat at night holding some fries, and he is looking away. Really close low angle skate photo, 12 mm fish eye lens, and it's a bit crooked with some visible film grain.

13:22 Then we're going to hit enter, and now it is starting to write the prompt for us using this formula. Then we copy this, we go back into Higgs Fields, we paste in the link, and then make sure you add in that reference image. I'm using now Pro for this as this works just as well. I just want to show you that this also works. And then we hit generate. And there we have it, Bruiser sitting outside on the curb.

13:45 It's looking quite cool. Like I quite like this one. Now, here's what the style can do for this image. So, for example, here I got the same image, but I changed up the style. So, this time I'm using a disposable camera wall flash style, and we got a slightly different image, but I might even like this one more. Here again, now we're going for the low quality camcorder style, and it's badly compressed.

14:04 So, I'm also using here the lowest resolution. So, I'm using the 1K resolution. Like on purpose, I'm making this look worse because imperfect looks more realistic in this case. You can do this for any type of style. So, here I made a iPhone selfie type of style. Again, you can see the prompt that I used beneath that to get that type of style. And knowing these different styles, you can make some very unique-looking images.

14:27 If you don't know again how to prompt this, I even put together this MD file for you. I'm giving away like all of the assets here. So, I'm hoping that this is useful to you. Please let me know in the comments if sharing these assets is useful to you because I put in a lot of extra time trying to make them universal so you can use them too even if you have no idea what you're doing.

14:46 One more thing I want to share with you, if you have made an image like this one for example, and you like the style of that, you can even ask your LLM to make an MD file, which is a skill file, for you. So, for example, here I ask, "Can you take the camera style based on the prompt that gave me this image and turn it into a reusable markdown file?" So, I gave it the prompt, I gave it the image, and then it made me this MD file.

15:07 That's how I make these type of things. You can do that too. If you have made something, you've no idea what you have done, you just put in your prompt, put in your image, and ask it to make that MD file for you so it's using that type of style next time you use your LLM. Now, creating the images is actually the easy part. The harder part is coming up with creative ideas.

15:28 And there are levels to this. Let me explain to you. We're going to talk about contrast because cool-looking images like these are like optimized for, I would say, social media. These do amazing on like Instagram. They all have contrast. You see something going on there that you look at and you're like, "What the hell? What the What is that?" And that's what I want to explain.

15:47 I will identify different type of levels of contrast. So, level zero is no contrast. It's nothing that is contrasting with the thing that's in the image. So, for example, it would be a robot in a factory. Like nothing here is standing out because this is the place, let's say, where it belongs. Level one is in an unexpected situation. So, here we have that same robot, but now we put him in an unexpected situation because he's sitting down next to a Tesla charging station.

16:14 And it almost looks like he had like one of those drunk nights out and he's about to just juiced up so he can continue with what he got. We could add in even more contrast by going over to level two, which is also adding in some unexpected styling. So, now we have the same robot sitting at the Tesla charger. He's giving us the finger, and he also has the I love Dan fully iced out type of robot.

16:36 This is culture. This is more like rapper attitude, and now he has a bit more personality. If we go one step further, we're going to add in feeling. Now, we are taking the example of the robot again. We're putting him in a field of flowers. Like you almost get a feeling of like, "Oh, I feel sorry for this robot. He didn't make it to that Tesla charger."

16:55 Those are the three levels of contrast explained with different images. And that is the thing that you have seen throughout all of the images and videos that I've shared throughout this video, which is adding in contrast. So, you have something cool that stands out, which can help you go viral. So, now that you understand contrast and how to make images, let's actually talk about how to make videos cuz that might be even more important.

17:18 If you want to achieve realism, you want to make it seem like someone actually filmed that stuff, like a human filmed it, not a robot. With Seid And Sue 2.5, we can achieve that by making it look like it filmed something like this. >> They didn't invite me to the party because apparently I filmed too much like a robot. So, he found a replacement. >> He's great, actually.

17:45 At least I film like a human. >> HEY, THAT'S MY CAMERA. >> SO, let me explain camera techniques that can be used to add in realism to your generations. First, we have handheld. This is where the camera has small shakes. It's just like someone is filming it using his phone and its natural shakes. The camera also responds to what's going on in the scene.

18:05 Then we have crash zoom. This is where the camera is suddenly zooming in fast and it's almost like the camera is surprised something is happening and wants to capture it really quickly. Then we have voyeuristic. It's almost as if the moment shouldn't even be captured. Then we have whip pan. It's a quick turn towards an unexpected event that makes it look like the camera is reacting to something.

18:30 >> Next up, we have walking backwards. Here, it actually feels like the camera operator is holding that camera and walking like backwards. You can see them like move unevenly with the steps. Another one I love which is not used a lot is blocking the frame. Here is where something happens and you block the frame by having someone like walk in front of it.

18:47 You really got to prompt for this. Like you got to think about this. And this is what you don't really see with AI. This is what you can see with real film things. So, that's why you got to think like a director. Another thing you can add in is exposure adjustments. This is what happens naturally when you go, for example, from inside to outside and the camera got to fix its exposure because it's all of a sudden super bright.

19:09 Then lastly, which is one of my favorite, which is focus field. Here's where the camera is struggling to adjust and focus. I will leave all of the prompts that I just shared with you into the asset folder, so you can go through this yourself at your own pace. The best thing you can do if you want to create like real realism is use these different techniques together, like combine them.

19:35 I put a few of them together and I got this like office style video for it. Shh. Now, if you're trying this out yourself and you're noticing that you're spending way too much credits, I've been there, then check out this video that's on the screen right now, because this video will teach you exactly how you can save credits on AI videos.