I Combined GPT Image 2 With Seedance 2.0- It's Suprisingly Good!

Added: 2026-05-07

[00:00:00]OpenAI is back. After a few slightly disappointing releases, they finally released something that is worth checking out. It is called GPT Image 2 and it is truly amazing when it comes to generating things with a lot of text or with a lot of context. Just take a look at these visuals. These weren't possible before with Nano Banana. Now, if we combine that with seed dance, you can make some pretty cool and interesting looking shots. Some idiot apparently prompted that a tornado will hit us. Is Paulie on the compile re channel?

[00:00:38]So today we're going to test it out. I will show you a few cool use cases and I will also show you the limitations of what's not possible and what is possible. I will also compare it a little bit to Nano Banana so you can truly see for yourself if it's worth checking this out. Now for this video we're using GPT image 2 through open art. The reason why we're doing this is because we have a bit more control over here. For example, we can choose the output which we have the option between all of these different outputs. If we compare it through Chetchup itself, you have less options here. Then the other thing that we have that we don't have in Chetchup itself is the quality. We have the option between low, medium, and high. The quality is not the same as the resolution. is not that this is 1K, 2K and 4K. Essentially, what it means is the quality of your images and also I think the time that it takes to generate it is different. For example, if we check the quality of low, it only cost us four credits. If we do high, it's 175 credits, which is a big incremental compared to the low. Now, what I'm still using is high in almost all of my generations. But if you want something fast and you just want to test out if your prompt works, then you can try it out with low, which will save you a lot of credits. If you want to follow along, I will leave the link to open art in the description down below. So, to explain the difference between low, medium, and high, I've generated the same prompt and the same reference image three times on all the different settings. So, here's the result. First, we have low. Keep in mind that the aspect ratio on all the three videos is exactly the same. The interesting thing here is that on the highest generation, it has the lowest like size. That's a bit weird here. But let's actually take a look at the image to see if we can see a difference between low, medium, high. So, this is low. We can see this doesn't quite look like Mr. Beast. We take a look at the contestants.

[00:02:29]It's not that great. Yeah, I don't know what happened here, but yeah, that's not good. Also, can we point out this thing right here? We can just see that it hasn't thought about it properly. If we take a look at the YouTube interface, it looks quite good. Although this button looks a bit too red in my opinion. Other than that, yeah, this is just not good.

[00:02:50]You can see that this is a fast generation. Then if we compare to medium, we can see that here. What the hell is this guy doing?

[00:03:01]I I still I like people saying that that Nano Banana is doomed forever. I I just still don't agree. I think this is good at some things, but it isn't the best.

[00:03:11]I'm also not saying that now is the best, but yeah, it still has some way to go. At least we have Mr. Be looking a bit better. I look a bit more realistic, I would say. We have no mess up on this thing. And in terms of text, buttons, all that looks pretty decent. Subscribe button is positioned a bit weird. It feels like this is coming from a playlist or something. Lastly, we have high. So, this took us the most amount of credits. Mr. Beast looks good. I would say I kind of look like I don't belong here. Like I've been photoshopped in there. If we come over to the text looks pretty good. This button looks slightly weird in my opinion. But yeah, again the people here it's just not good with the background. It it feels like it has used all the compute of generating realistic images of me and Mr. Beast. So yeah, that's the difference between low, medium, and high. For reference, I ran the same prompt through Nanomana Pro.

[00:03:59]And here at least the people seem to be better. It also feels like it's way too oversaturated. It kind of looks like a thumbnail from Mr. Beast. Yeah, this is not bad, I would say. Then I also did it on Nano Manenna 2 here. I would say the interface of YouTube looks a bit more realistic. But I have to be honest here.

[00:04:18]If we take a look at the text, then we can see that some of these comments just don't make sense. This was the circuit is stretch. I don't even know what language that is. So yeah, that's unfortunate. I would say Nanomina 2 understands the context a bit better than Nanomana Pro. So yeah, all in all, I'm not too convinced by that example that I just shown you with the high compute. The only thing I do like is how good the text looks. The background just doesn't work. But hey, if you combine the two, you can make it work pretty well. The second thing that I want to show you, which CHP is very good at right now, is generating images with text. For example, I generated this screenshot of me talking to Dario. And honestly, I'm kind of frightened here because the amount of people faking text, faking Snapchats, faking any type of screenshot is going to be insane. The amount of fraud that's going to happen with this, I hope if you're a bad person, please don't use this to your advantage. Yeah, I I hope you only use this for pranks. But just look at this.

[00:05:17]The text is good. The today is good with the time. We even have the time when I took the screenshot. It all makes sense.

[00:05:23]And I just gave it a simple prompt. I just said, "Create a screenshot of an iMessage chat with Dario Amodai. It's a funny interaction between him and Danke.

[00:05:30]He mentions Dan made your legend." And because of that, you get unlimited Claw Code credits for the rest of your life.

[00:05:36]I wish, right? Make the rest of the interaction funny. So, I just let it fill in the gaps. And it came up with this. Now, this is basic. You can do way more advanced stuff with this. For example, what I did here is a bit more tricky. What I did here is I basically went over to YouTube. I took one of my videos and I literally took a screenshot just like this and I put it into Nano Manana. So I took a screenshot like this, put it inside of Open Art. Let me not do that right now, but I took it in here as a reference and I set a photograph of someone photographing their PC monitor while they have YouTube opened in that YouTube screen. They're watching a guy from image one. That's his reference. And it needs to be candid low quality. I don't want to see the phone from the image. I tried it a few times. Sometimes I had him hold the phone, which is not the point that I want to do. It's like he's taking an image of his PC screen. Like if I were to take an image like this from my screen that I have in front of me right now. It made this Snapchat type image which I wanted to have it done. Like we even have the text which looks exactly right. Wrote this guy's video is insane.

[00:06:34]Definitely the goat of AI for real. I mean I wish someone sent me that. Yeah.

[00:06:39]But all in all that's crazy. Then I tried something else. I tried combining this with Cance. And let me show you this. So here I went over to video and then I put it inside of Sea Dance and I gave it a prompt and I wanted the person to talk about it.

[00:06:52]>> Yo, this dude is GG goat. Everyone should watch this video. His name is Dan.

[00:06:57]>> Unfortunately, it didn't work. Like the voice over worked, but me lip syncing to that doesn't make sense. I tried even prompting it. Like the person recording him is speaking, not the person on the screen. Tried it like three different times. Got three different results. Yo, this dude is the goat. Everyone should watch this video. His name is Dan. Now, it's not good yet, but if you can make it work, then you can make some pretty interesting shots of it. It's pretty cool that you can have chip tea combined with seed dance and make it look like it's actually real. It's also scary at the same time. I keep reminding myself like don't do bad stuff with this. Not that I intend to, but you get what I'm saying, right? Like, you can fake testimonials. You can pretend like you ordered something. Yeah. I know I'm not going to give you any ideas here. Let's move on to the next thing here. Here we got one more example of me having a chat with Mark Zuckerberg. Like this in Nano Banana wouldn't look as good as this. I know of course you can Photoshop this easily, but just how quickly you can do something like this with just a simple prompt is pretty cool. Another example that I've seen going viral online is creating a poster with a detailed color analysis of what your best features are and what the type of color you need to wear. I tried it out just before this video. put in this selfie and then apparently my season is soft autumn. Go crazy with this. Whatever you like, whatever you want to hear it, AI will tell you that. I'm taking this with a grain of salt. Like it looks good. But yeah, if I'm going to be honest, I still think I can rock blue. I still think I can rock black. Pink, maybe not. But yeah, this is pretty interesting. The other thing that I saw which I had a lot of fun with was this one. I tried this a couple times with the best hairstyles.

[00:08:38]So, I should have done this before I went to the barber. Apparently, this is my best hairstyle or this one. I tried this one more time with this one. So, here we also have my face shape. We have the styles that look best for me.

[00:08:52]I I kind of get why I can't do the slick down one. Jokes aside, yeah, this could be this could be pretty useful for when you go to the barber. I also tried animating this with seed dance. So, here I said animate this image. Make each person turn their head so they show their hairstyle. This is just being done with Cense 2.0. I did it in 1080p, but just don't waste your credits. Just use 720p for this generation. Basically, what we have here, which I'm not sure if it all makes sense, the side angle of the head all pretty much looks the same.

[00:09:22]That's the only limitation I'm seeing here. So, apart from the top side, like the the low fade, I mean, that's not what I'm rocking right now. Like, maybe should have asked my barber for that, but I don't think it look good on me.

[00:09:34]But yeah, that's some inspiration for you right there. Another thing that I've seen going viral is creating images of yourself in a movie poster. For example, this one with Will Smith here. I made this one. Create an artistic movie poster with me, Dan Keeft, and Will Smith in the lead for a true detective movie. Try to add the double exposure effect that a true detective is known for. Both men have a serious look to their face. So, I've added in my name so it can add that in the poster as well.

[00:10:00]And honestly, this looks impressive.

[00:10:02]Like me coming from a social media background where we used to pay people good amount of money to make poses like this. It's kind of sad and exciting at the same time that you can now do that with AI because the question is like how does it come up with that stuff?

[00:10:18]Probably being trained on actual artist work. Then again, can we avoid this? No, there I don't think there's any stopping to it. So yeah, I tried this a few times with different results. Like for example, here we have me in the GTA movie and this looks really cool. Like not saying that you cannot do this with Nana Banana. In fact, I tried it here with Nana Banana. I got this result which in my opinion also looks quite good but not as cool as the the other one. Also, the text is completely screwed. Like the the names and even the words in the bottom right here don't make sense. Try it one more time using Nano Banana Pro. I kind of love this one, but the text again is a bit screwed. But still, the style of this one is quite cool. But this one gives me more of the newest GTA 6 type of vibe.

[00:11:02]Here are four more movie posters that I made with GPT image 2. All I did was ask create a movie poster for like Fortnite, Roblox, God of War, or GTA. Pretty sick, right? Now, let me show you another example where we combine this with Cance 2 because in my opinion, that's the most exciting thing that I like using image generators for to bring it to life in a video model. I've noticed that for like lowquality realistic type of videos, GPT image 2 works great. For example, I made these six totally different type of I would call them like screenshot vibe lowquality images of pretty interesting scenes and I used a pretty simple format here. So, if you analyze my prompt, let's do this one from the bear. We can see a shaky raw shot from a hiker's chestmounted action camera. A massive brown beer is standing just a few feet away on a narrow forest trail, staring directly at the camera. The hiker's hands are visibly in the foreground. The photo is full and heavy motion blur, digital grain, and looks like a terrifying lucky escape without any text overlays. If I were to break this down for you, I'm using a combination of a camera style. So, that's going to be a shaky raw shot from a hiker's chest mounted action camera. That's the camera style that I'm using here. Then I'm using subject and action. So a massive brown beer. That's the subject. Then the action. This kind of goes hand inhand with the subject is standing just a few feet away on a narrow forest trail staring directly into the camera. Oh, and then also the hiker's hands are visible in the foreground that also kind of belongs to the action. It's just a few more details. Then for the lighting, we don't really have any lighting here, but usually I would say something like lighting here. In this case, I'm using the grit, which is photos full of heavy motion blur and has digital grain and it looks terrifying. Lucky escape. And then the last thing I'm saying is no text overlays because sometimes I was getting some text overlays or it took like those if you have a screenshot, it had like text in there of like the date that it was recorded or whatever. So that's what I did here in these images. If I then put this into seed dance, we get something like this.

[00:13:28]That's pretty terrifying. Right now, the way I made this is as followed, and it's easier than you think. In fact, I didn't use a prompt at all. Like, Cance is pretty good at just understanding the context and knows what needs to happen.

[00:13:40]So, no prompting being used. I tried this with a few of the other shots here as well.

[00:13:50]Can Can someone pray for my dude over here? Like what's going on with him?

[00:13:56]Man's trying to choke himself. I found that it works so well if we combine these like realistic lowquality type of shots with seance.

[00:14:18]I didn't even ask it for the slow motion and it's not even possible if you were to do it like a realistic shot, but still it looks cool. This one too also.

[00:14:35]So yeah, if you want to risk it sometimes you don't have to prompt if you just want to see what CS makes of it. But of course, you need to have some kind of clear thing going on where Cance has enough information from the image alone where it can make something looking good. The next thing I wanted to try out is me doing a weather report.

[00:14:52]So, I have this image of me. I asked generate an image of a screenshot of a live broadcast of a weather reporter reporting the news during a huge storm.

[00:15:01]But he is showing signs of discomfort.

[00:15:03]The camera is a bit low quality. Make it as realistic as possible. Put him in a yellow raincoat. So, that is all I did.

[00:15:10]I said output 16x9 quality high generate. That gave me this result. If we now put this image into seedance, then we get this. Some idiot apparently prompted that a tornado will hit us. Is Paulie on the copter? Is that piles of re channel?

[00:15:31]>> Besides the American accent though, I I love this video. Like the way to fix this was to probably do like audio to video and me saying this myself and then putting it in there with this reference.

[00:15:43]But yeah, just for time sake trying it like this. Pretty cool. Good result. And like it keeps the text here consistent without me having to prompt anything on that. I just kept it really really simple here. The next thing I want to look at is translation. How good is GPT image 2 with translating your images?

[00:16:00]So, I made this image of a Chinese takeout menu using this huge prompt and it gave me this a lot of text going on and I made this with GT image 2 because then we know that it doesn't mess up with text. We got this in English right now. Let's translate this to another language. So, here I did it in French and to be honest it looks pretty good.

[00:16:21]As far as I know my French, which I know it a little bit because I didn't pay that much attention in school. The words look right. Like I know this is beef. I know this is whiff. I know this is greens. Seems about right. Then I translated this French to Dutch. And this is the language I do know. It is pretty correct. It it does look like some of the translations are a bit too like done in a weird way. Like some words in in Dutch, like especially on a Chinese takeout menu, we would put them in English. But overall, I understand all of it. It it looks clear to me.

[00:16:57]There are no mess ups in the translation. Here I have another example of a few parking signs. So I made this using GPT image 2 again. And the thing that I dislike when I translated this to French is it made it way brighter. Like that was not my intention. Like maybe I should have said keep it in the original colors in original style. Then I did the same to Dutch. And here I would say though I like this means don't park but it it doesn't make sense. It's like it's it's I mean it's technically it's correct but I would say differently. Other than that it's it's really good. Another thing that's pretty good to play around with with sometimes doesn't work you might get a few generation fields here is creating images of game play of any type of game and then bringing them alive into seed dance. So for example I use this prompt create an image of gameplay of PlayStation one version of the last of us. I want a screenshot from the game play that got me this image. Then I put that same image into Cance 2 and I basically set gameplay of The Last of Us while this character is just walking around fighting characters from the game. This is what it made.

[00:18:11]Okay, besides that, this guy is just running through this concrete block there, which might be a game bug.

[00:18:22]the sound and the movement look like a video game. So, that is my first look on how you can use GT Image 2 together with Seance. If I've missed any of your favorite prompts, let me know. I will do a follow-up video on this. And if you want to try this out yourself, then use the link in description down below. It will help support the channel. And lastly, if you want to see all the prompts that I've used throughout this video, then make sure that you go to my school community in the YouTube resources. You can find all of the prompts and all of the files that I use for each and every video that I post.

[00:18:55]Click the video that's on the screen right now if you want to learn how to create cinematic AI films using Cadence