
This is the first time an AI model has impressed me. In today’s video, we are going to learn how to create VFX sequences like this, how to create videos which you can use in your client projects, and how to become a director and create multi-shot sequences with a single prompt. We’ll also learn how to create Meta ads and commercials that actually look good. A VFX sequence like that used to require months of work and lakhs of budget, but now you can generate that in a few seconds. Six years in this industry, and this is the first time an AI tool has really impressed me. Today I’m going to show you the step-by-step AI generation method that you can use to create realistic AI videos
First, let’s start with why there is so much hype around this model. So this is Seedance 2.0, and it is a multimodal AI video generator. You can feed it text, images, video clips, and audio files all at the same time—up to nine images, three videos, and three audio files in one generation. Now there is so much hype around it because it’s getting a few things right. First is camera movement: you can specify exactly how the camera moves. But the real step up is that it builds coherent multi-shot sequences with the same character, same world, and same energy across every single shot, which means you type a single prompt and you’ll get the complete sequence ready. That’s the thing which has broken every other model till now.
Now let’s talk about what you can actually do with this model. First, product commercials. Let’s say you’re selling something on Amazon and you have just got your product images, nothing else. You can drop those photos in as reference, write what you want the ad to feel like, and it spits out a full commercial which has transitions, cinematic effects, time-lapses, freeze frames, you name it. Second, it’s really good at fight scenes. Every other model feels like AI slop because of the amount of physics involved. The problem was that punches go through the body, reactions don’t match, and physics fall apart. But it’s completely different with Seedance 2.0. With this one, punches land, the body reacts correctly, and the camera cuts to a different angle while keeping the character consistent, especially for action content. Nothing comes close to this. Next, this model is really good for character animations, 2D and 3D animation, and also to convert manga to animation, which means you can upload a comic or manga page and it will become a complete animated scene. Next, it can be used for dialogues and performances. You can write a scene with specific emotions and it performs them—it can do whispered lines, screaming, crying, and even lip sync in eight languages. And if you have some cinematography knowledge, you can actually feed it different camera angles and create a complete sequence. And last, this is the most viral format coming out of this model, which is a cinematic suit-up. Start the scene with something boring, like someone in the market or someone driving, then something breaks the calm. A different character comes into the scene, and then the transformation happens—show how, not just what happened. That’s the format, and you can build the entire thing with one reference photo.
So here’s the full step-by-step process. Now most people get this completely wrong. They type a simple prompt themselves and then hope the model will do the rest of the work and create a good output. But AI photo and video generation does not work like that, and after that the output looks very mid and they blame the tool for that. But the tool is not the problem, your input is; AI video is only as good as what you feed it. So let me show you exactly how to do this. First, open two things together: OpenArt and your preferred LLM, so that could be ChatGPT or Claude—for me, I’ll go with Claude. Once you open OpenArt, then you can go to video generation here and select the Seedance 2.0 model. Once you do that, you’ll see you have two options: you can add start and end frames, and also you can do text with reference
Now before you touch anything else, you need to know what you are building. The more decisions you make before generating, the less the model has to guess, and every time the model guesses, it will get it wrong. So first we’ll start by creating a character. To do that, click on “Character and World,” and here you can create a character. I have already created two of mine, and you’ll see I just uploaded multiple images of me—one front-facing and a couple of side angles. This will lock in a character. Next step: I want to create a samurai scene of me fighting people. Now I went to Claude and I asked it to create the prompt for Nano Banana. A good image generation prompt needs specific art direction, lighting, and also camera language. Claude pulls from thousands of visual references you would not think to include, and the output is just better every time. Once this prompt is ready, I’ll copy-paste it into image generation in OpenArt, select my character, and click on generate. Nano Banana generated these frames. Once the suit was ready, it was time to create the environment and the people who I’m fighting. I went back to Claude, uploaded this image, and asked it to create the prompt to generate a similar environment and also the people that I’m fighting. Once done, this was the image that I got, and now I’ll upload all these images again on Claude. After uploading all these, I’ll ask it to generate a very detailed Seedance 2.0 prompt for an epic fight sequence. So I’ll copy-paste this prompt, go to OpenArt video generation, select the Seedance 2.0 model, choose text with reference, upload the prompt, select images 1 to 4, upload those photos, and then click on generate.
This is the video that is generated. Just in case you get an error or you did not get the right output, you can go back to Claude, tell it what the error was or what’s glitching in the video, and then try again. AI video generation is all about iteration. The more you try, the better the output will get, and you’ll also know what to do. Now I have been testing this for a few days, and usually if you get an error, there are a few things going wrong: one could be the prompt is too long, and second, you can check the generation failed message on the bottom to see what the exact error was. I have realized these AI models work best when they are working with real footage. So I recorded a video of myself filming my laptop with a creature inside my laptop, exported the end frame, uploaded it to OpenArt as the first frame, and added the prompt that I want to hold the creature and grab it out of the laptop. After that, I exported that video, imported it into Premiere, and merged them.
The best part about OpenArt is that they already have pre-made templates. For example, you can click on a template to see the prompt it’s working on, click “Use,” replace the default image with your character, and click generate. I really like how detailed my face is in this, plus it looks like it’s been shot by a real camera—the hand is blurred and my face is focused, which gives it more depth. Along with that, the music and sound effects are completely on point. The best part is prompting it once after uploading high-quality images of yourself; if you’re using a template, this works really fine. Next, since the samurai suit setup was already ready, I uploaded the same image again and selected another template. Once your setup and character are locked, it does a really good job.
Now here are other use cases that you can actually use in real-life client projects. I’ll provide this complete PDF to you along with all the prompts and all the images that I used so that you can try it for yourself. Eighty percent of people who regularly watch the channel do not subscribe, so if you can please take a few seconds and subscribe, that will help us a lot. That being said, the next use case is product commercials. The process is again very similar: you upload product photos to Claude and ask it to give you multiple frames of the product. For one test, we tried the product around a snake, generated a five-shot prompt, and copy-pasted it. Next, we tried the same with a chips packet—for one-click generation, these are really good outputs that can be directly published with just minor tweaks. I tried another commercial design using headphone images and also uploaded a callout image to see how good Seedance 2.0 is with tracking. Next, this is the best model to do anime or manga sequences. Seedance handles anime style better than most models and provides very high-quality animations.
After the fight sequence, I spent a lot of time testing what else it can do, and here’s what I found. Animated UI elements floating in 3D space work really well—it can generate clean holographic panels, data updating in real time, and perfect background assets for tech client videos. The glass surface rendering and volumetric lights are cinematic, and once you combine this with real video, like I did with an Asus laptop, the output looks really good. Next, particle effects and energy animations are one of the best use cases; the spiral vortex and energy burst outputs are stunning and perfect for brand intros and music video intros. Next, data visualization animations look really good, though don’t expect precise numbers or accurate graphs—use it for visual aesthetics rather than actual data, or composite real numbers in After Effects. Next, product renders with animated callouts are genuinely sellable; the headphone commercials with callout lines came out clean, so you can upload your client’s product image and make a real deliverable. Finally, the audio design is really good—you can be very specific about what kind of sound effects and music you want, and it will sync all of that with the movements.
I have been testing a lot of AI tools for some time, and every single one has broken on something, but Seedance 2.0 is hitting 90% accuracy across multiple scene types. If you know how to prompt and direct, this can be really useful. I used AI video generation in my last project, and I generated 40 to 50 videos to get maybe one or two usable outputs—that’s just how AI video has worked, but this gives us promise of a better future. People who figure this out and test these models right now will be very difficult to compete with in a couple of years. Once you understand how to use these AI video models in your projects, you will be unstoppable. To try it, you can use the link in the description, go to OpenArt, and try the Seedance 2.0 model. The barrier to making high-quality video content has been dropping every single year—first it was cameras getting cheaper, then editing software, and now it’s this. We are already seeing multiple hybrid creators combining traditional editing skills with AI techniques.