Seedance 2.0 + Video Reference = Fixes Continuity Issues

本站添加: 2026-05-08

[00:00:00]Nobody move. Just stay where you are.

[00:00:10]Drop the weapon. Hands where we can see him.

[00:00:15]Hands up. Come out slowly.

[00:00:24]Keep your hands up. Walk forward slowly.

[00:00:29]Don't move. Stay where you are.

[00:00:33]>> What the hell is that?

[00:00:36]>> Hold your fire. Hold your fire. Back up.

[00:00:39]Back up. What is he turning into? Take cover. Move.

[00:00:47]>> One of the biggest problems with any AI model, including Cance 2.0, was when you generate your first video, what happens after that? When you generate the next video, there are a lot of continuity issues. The scene might look different, the person might look different, the music might be different, the sound effects may not match. And all these continuity issues used to make creating longer scenes a real task. But now with video references in Cance 2.0, this has become so easy. Let me show you the scene that you just saw, it was so easy to create that. First of all, for the first scene that you saw, that was a 15-second long video. And this was your usual process because the first video is never the problem. So, I'd gone to Chad GPT, explained it the script about, you know, the cops coming in, catching him, and then he flies away like a typical transformation scene. And this was the first clip prompt that I got. So, this is no issue at all, right? We can do this with any video model. So, I'm using hick wheel. The link is there in the description. I really feel Hicksfield is the best way to use Cance 2.0 simply because you hardly ever get unsuccessful generations. So once I was over here, I selected Cance 2.0. I wanted to use this guy. So for the first clip, I have uploaded his uh reference image here and then I basically took this prompt and copied it. And for the settings, I had gone in for 21x9 aspect ratio, 15 seconds, and a 720p resolution. Do note that you can now generate on 1080p also, but my point is to just show you how things work. And then I was able to generate this video.

[00:02:30]>> Nobody move. Just stay where you are.

[00:02:40]>> Drop the weapon. Hands where we can see them.

[00:02:44]So no issues till now. This is Cance and it gives quality results. Now the problem is when you want to continue to the over to the next scene. So when it comes to the prompt again there's no problem because you can go back to chat GPT explain what you want and it will give you a detailed timestamp prompt like here clip two you can see. So I've got the prompt but this is not the problem because if I was to use this prompt only and again just use his reference picture just like I did for the first video. Who knows what you're going to get, right? But this is where video references come into place. So here along with the picture, I can also go to this part and then I can go under video generations. Now because I have already generated this first video, this is going to reflect right here. So this was that time the first video. So I'm going to click on this. And at this particular point, Hixield will ask you if you want to trim this video reference. So first of all, what is a video reference? Well, just like you give some kind of context references to sea dance, this is another way of doing that. So, you're basically telling Cance, listen, this was the first scene that we created. Yes, the prompt is different, but remember this is a continuation of the scene. So, Cance goes through this video. It sees exactly what is happening. It sees the uh notices the sound effects, the music, the kind of color grading, everything, whatever is happening. There is no better way to tell Cance what the context is than simply by showing it what it generated. Right? So then if you wanted to give only a part of the scene as the context, you can trim it down also. But let's say you want to give the whole scene, no problem. So we can just say save and after some time Pixel will check for eligibility and it will come right here. All right. And just like that you can see that we've got our video reference here. This is the whole video playing here. And of course, I've again used the reference image. And you can just tag this whenever you copy your clip to prompt. So you don't have to, by the way, one thing I've noticed is you don't have to even tag this right at this particular place, unless and until it's really important. You can even just before you start writing the prompt, you can just tag both the video and the image. And that way, you know, you don't have to worry about where exactly are you tagging it. And what's going to happen is you're basically going to get a video like this.

[00:05:03]>> Drop the weapon. Hands up. Come out slowly.

[00:05:13]>> Keep your hands up. Walk forward slowly.

[00:05:17]You can see if you compare that to the first video that we had got that exactly looks like a continuation of the first scene. That's because Seance knows everything right now. The only thing that I have noticed is whenever you use a video reference, most of the time seed dance will start the next video that you're generating. If you're using the video reference from the earlier video from the ending frame of the last video that's not a problem because at the end I'll be showing you when you do put this in the editing software, you just have to cut out that initial part. That's all. Okay. [snorts] But similarly, I used this second video that we had generated as the video reference for the third generation. got the prompt from Chad GPT described the transformation of this person flying away and all these things and this was the result.

[00:06:02]>> Keep your hands up. Walk forward slowly.

[00:06:07]>> Don't move. Stay where you are.

[00:06:11]>> What the hell is that?

[00:06:14]>> Hold your fire. Hold your fire. Back up.

[00:06:17]Back up. What is he turning into? Take cover. Move.

[00:06:25]Now the good part about this is that no longer are we relying on some kind of screenshots or some other references which used to be a real task. Now because we are using video references our focus completely is on the result itself because we don't have to worry about what we're uploading here. Right?

[00:06:44]And this is the fun part. So it's not like yes the video references will always give you the best result because that will depend on the prompt also.

[00:06:52]That is different from being frustrated with continue to issues because that's a separate game altogether. At least now we're only focusing on the result. For example, for the final scene, yes, it was still challenging. So I did have to generate around three generations because for example, I was not liking the transformation. Like this was one of the transformations where he suddenly turned into this jet and I did not like it. But this is not a continuity issue.

[00:07:15]This is just a quality issue and you can just, you know, change your prompt a bit and this is a lot of fun to do. But this is not the same as being frustrated by those issues that we used to face earlier. And then once you've got your videos, let me just show you how easy this is to stitch together in something like Premiere Pro. So you can see this was the first clip and this was the second clip. Right now I've just intentionally left a gap there just to show you this. So you can see right this ended right here because we were using a video reference. The next video also started from this. Now, this is something you still want to avoid even though it gets it pretty close, but you don't want exactly the same frame because there will be an evident flicker between the two. So, what you can do is in the next clip, right where the scene changes, you can just cut it and then join this together. So, it then seems very seamless. But the main solution here is that the other things that are happening inside this clip, they all are similar to what's happening here because this was the reference for this. And talking about efficiency, Hicksfield have released finally a nodebased structure to create these videos in a seamless way called canvas. However, this particular thing, so if I was to click, I've already tried it here otherwise you would start with create new canvas. But let me just show you what I mean here. So this is same as creating those videos is just a node based node based system. So you know you can use it's repeatable basically. So you can see here my reference image then I'd use the node for the video generation. Now what I was hoping for was that you know if I just add another node here. So let's say if I was to go to video generator again select C dance here and then connect this video generation the first scene to this then it acts as a video reference currently that is not happening. So the video reference part that we saw before where it shows up in the references box right now when you do it using hickfield canvas it shows an error. So I'm trying to see what is happening there. But once I do find a solution to this, this will even make things much faster and great for film making. So I'll probably work on this and if I do find a solution, I'll definitely make another video on it. In case this one helped you out, do give it a like. And for more AI editing tutorials like this one, make sure you subscribe.