AI vs VFX: Can Seedance 2.5 Beat VFX?

Every shot in this workflow was one prompt in Seedance 2.5 without 3D, compositing or render farm. Steal the prompts above, swap in your own hero, and see how close you can get.

How To Make Realistic VFX Shots With Seedance 2.5

Seedance 2.5 just went live on TAVOA AI - it's the most anticipated model of the year, every single shot below was made with it, and there's a special launch offer running while it lasts.We ran a challenge with Erik - the VFX artist behind some of the most viral VFX videos on the internet - to see if AI can actually beat VFX. Same shots, built two ways: he uses the full VFX pipeline - 3D tracking, simulation, compositing - and we use only AI. But the real question behind the challenge isn't about beating anyone. It's this: can AI help VFX artists do their work faster and easier?First, watch what we had to recreate. Erik traveled to Iceland and filmed real footage around a cave - a drone shot toward a waterfall and a few shots inside the cave itself:

The brief: put our host into that cave, make him jump off the cliff onto a dragon, and fly through the waterfall - frame for frame. And then take it one step further: a fully AI-generated cinematic that would cost studios at least $50,000 the traditional way. Here's how we did it, step by step.

Step 1 - The Assets

One photo against a wall is enough.

Upload it to Claude, ask for a character sheet on a grey background - full-body front and back plus a close-up - and generate it with Nano Banana 2, the best model at matching faces. And here's the tip that saves you real credits: always generate character sheets on a plain grey background. Keep this workflow in mind, because it runs the whole project: Claude writes the prompts, TAVOA AI makes the content.Save this image as an element and name it @adil-hiking

For the dragon, switch the model: Seedream 5.0 is the best at fantasy creatures. One key tip for a creature with this much detail: generate two close-ups - mouth open and mouth closed. That keeps the face from glitching between expressions.

In traditional VFX, this asset alone is a project: you'd sculpt the dragon, do the retopology, unwrap the UVs so the texture fits, then rig it before you could even animate - and only pros with years of experience can pull that off. Buying a model costs real money, and you'd still have to animate it yourself. We made ours in one minute for 40 cents.

Step 2 - The Video-to-Video Swap

To recreate Erik's footage we use video-to-video: upload the original, make the edits with a prompt. First, cut the source into clips - and if any clip is shorter than four seconds, freeze-frame the last shot: Seedance needs at least 4 seconds of video input for v2v. When you generate, set the duration to four seconds to match the reference - anything longer and the AI starts guessing.This prompt skill breaks down camera angles, lighting, and movement so Seedance knows exactly what to do. Download it for free here

The AI matched the lighting and contrast to the environment automatically - pause mid-action and the frame is still crisp. In traditional VFX this shot needs a compositor: keying (curly hair is a nightmare - every stray hair cleaned by hand), then despilling the green reflection off the skin, then compositing the actor into the scene and matching light, shadows and film grain. That's hours, even for a pro - and it's exactly the step AI now does under the hood.

Step 3 - The Jump

The dragon-catch shot failed across all four batches - and that's the lesson: there's no actual jump moment in the source, so v2v turns it into slop. If the same problem shows up in all four batches, the issue is the prompt or the source material. Batching further just burns credits. The fix: switch to image-to-video - take a screenshot of the location, use it as the starting frame, and stitch the two shots in the edit. Notice the empty-frame pause written into the prompt: 1.5 seconds of nothing but waterfall - that's your cut point.

To pull off this jump the traditional way, you'd build a full CG double: cloth simulation, physics, body-mass and trajectory calculations - all for a two-second shot, plus rotoscoping the midground. It's a huge grind, and it's exactly the kind of work pros like Erik do. And CG doubles are where the viewer's eye catches the fake: there are only two ways a person goes off a cliff - you step off vertically, feet first, or you push off and dive. When a digital double tips over in a stiff "soldier" pose and slides down at one constant speed, with nothing moving in the clothes - that's not a person falling, that's an asset being dragged from point A to point B. Your eye catches it even if you can't name it.

Step 4 - The Burst

In traditional VFX, realistic creature movement has its own anatomy you have to recreate: build a 3D scene, match the lighting, animate every part of the rig - down to the scales. Just animating the dragon could take weeks, and the standard 3D software alone costs up to $3,000 a year. Here it's a rewrite and a generation.

Step 5

The fix is a size-ref frame: one image that locks your exact scale on the dragon, saved as its own asset and attached to every scene. Build it by combining the dragon sheet, your sheet and a screenshot from a good previous shot:

Then regenerate the flyby with the scale law attached - and keep the camera shake and lens droplets the prompt asks for at the end. That's what makes it read like a real drone:

Part 2 - The $50,000 Cinematic

No real footage at all this time - the whole scene is generated. The pipeline is the same - images first, then video - so here are just the new tricks, scene by scene.

Generate with GPT Image 2, then learn to spot slop: a warped logo, a blue-ish color shift. For small fixes don't write a big prompt - switch to Nano Banana 2 (the model that holds inputs best) and write one line:

The new sheet - and the face-lock trick:

And the face-lock trick: crop the head out of the full-body frames, so the AI has exactly one place to pull the face from - the close-up

Remember the size-fit tip from Part 1? Let's make the AI's job easier and merge the images into one. I took a frame from the dragon's original sheet and my new character sheet, and combined them with Nano Banana 2.

Locations

Scene 2

How to get real emotion with no video reference: describe the face in detail - what the character is going through, what he says, how his voice sounds, what his hands are doing. The swallow, the twitching fingers, the clenched jaw, the shaken-out hands, the exhale, the sprint. Seedance handles the light, and it comes out natural with zero compositing.

Scene 3

This time the whole beat runs in one continuous shot - and here's the voice tip: Seedance locks the voice along with the appearance in the character sheet, so the scream stays consistent across generations without describing it every time. The camera chases the run, stops dead at the cliff lip, and the one-second empty frame builds the tension before the dragon erupts.

Scene 4

Scene 5

For high-speed shots, put the camera language right in the request: a 180-degree camera orbit, speed shake, and lens droplets for realism. Dynamic without turning the subject into blurry slop.

Scene 6

Clouds are a huge pain in traditional VFX - millions of volumetric particles, hundreds of gigabytes, render farms crunching for hours to produce one second of footage - and here they all have to react to a dragon flying through them. The secret sauce is motion physics: always specify physical movement in your prompt, and on high-speed shots let the AI improvise background detail as long as it doesn't break the story.

We hope these hacks make working with 3D and VFX easier and more accessible — and if you've been wanting to create complex shots like these but never had the technical background, now you can just try it with AI. I've left all the prompts in the description, so take them

Your First VFX Shot Is One Prompt Away

by TAVOA AI

Discover more

AI VFX Just Got Insane — Seedance 2.0 in 4K Turns Real Footage Into Movies