Kling vs Seedance: I gave the same prompt to 8 video models

I had eight different AI video models make the trailer for my game, Operation Tea. Every one got the same prompt, the same five images from my game and the same length. First I explain how these models work, then the results.

1. WHY I RAN THIS TEST

There are a lot of video models on the market right now. I wanted to measure what kind of performance we get when we give the same prompt to eight different models. I also thought: we pick up models and use them, fine. But if we have even a little idea of how the model on the other side works, we can write better prompts.

Making video is hard and takes a lot of time. However much I explain to Claude Code, it gets me only so far, then it gets stuck. In this video I wanted to save you the trouble of trying every model one by one. The video was made in collaboration with Higgsfield; all eight models sit in the same model picker on Higgsfield, switching models is one click and the prompt doesn’t change at all.

The contenders: Kling 4.0, Seedance 2.5, Seedance 2.0, Grok Imagine 1.5, MiniMax H3, Gemini Omni 1.1, FLUX 3 and Wan 3.0.

2. HOW ARE VIDEO MODELS TRAINED?

To train a video model, you give it millions of videos. Next to each video there’s a text describing what happens in it. During training the model reads the two together. (An example of such a video-and-caption dataset: Panda-70M.)

The interesting part: those texts are usually not written by people. Writing captions by hand for millions of videos would take a lot of people a lot of time. Instead, another AI model prepares the dataset.

What this means for you: when you describe your prompt the way a person would (“the car should go like this, there should be a girl”), the system sometimes first translates it into a form the video model understands, then sends it on. Because the training texts were written by another model. So your prompt either gets rewritten behind the scenes or doesn’t give you the output you wanted.

3. HOW DOES THE MODEL READ THE PROMPT?

Your prompt first goes through a language model, called a text encoder (T5 and CLIP are examples). Your sentence is turned into a very long list of numbers called an embedding. The model doesn’t read the words, it reads these numbers. Whether you write in Chinese or Turkish, in the end everything becomes numbers.

Video isn’t processed as a sequence of photos either, but as pixels flowing through time. Just as language models split text into tokens, video models split the video into small three-dimensional cubes, across both width and height (space) and frames (time) (ViViT). Then those cubes are compressed, like zipping a file, into a mathematical space called latent space (Latent Diffusion). That makes the data the model works with much smaller; when it’s done, the result is unpacked back into video.

4. HOW DOES DIFFUSION WORK?

In training, a real video is taken, say a video of a dog. In a step called forward diffusion, random noise is added to it step by step until only a grainy, unreadable screen is left. Then reverse diffusion: starting from that grainy screen, the model learns how to remove the noise so that a meaningful video comes out (DDPM). When you give a prompt, the model starts from scratch, from a grainy screen, and cleans it step by step into a video.

In models from a few years ago, objects fell apart after the second or third second: ten fingers instead of five, backgrounds twisting into strange shapes. What solves this in the newer ones is the diffusion transformer (DiT): the diffusion structure combined with the transformer architecture. The transformer has two jobs here:

• Cross-attention: it matches the words you write, say “a walking cat”, with the pixels in the video (cross-attention maps).
• Temporal consistency: the model knows the object in frame one and the object in frame fifty are the same thing. If a car drives away from the camera, it works out what the back of the car should look like.

5. WHY IS MAKING VIDEO HARD?

The high-quality, non-slop videos you see online probably came out of a lot of effort and attempts. I don’t think a great result from a single prompt on the first try exists. People get really good results after a detailed prompt, good images, a good plan and a good character sheet.

I did a few tests before this video. In one, I gave the model a single frame from my game, a street. The model made up everything outside the frame, the rest of the street, the hidden part of the car; three seconds later we were on a completely different street. That’s where guessing kicks in. When I gave more images and wrote a more detailed prompt, the model stopped guessing and started following the images I gave it.

6. HOW I SENT IT TO 8 MODELS ON HIGGSFIELD

I ran all the models on the Higgsfield site, on screen. Every one got the same prompt, the same five images (a start frame, the front of the tea house, the inside of the tea house, the characters, the cars) and the same length, 10 seconds. Ten seconds isn’t enough for a trailer, but it’s enough to compare models. I wrote the prompt by explaining my idea to Claude and working on it together.

I scored on five things:

• How closely did it follow the prompt?
• How natural is the motion?
• Did the characters and cars stay the same from start to finish?
• How is the audio?
• And most importantly: does it look like AI slop?

The prompt I pasted unchanged into all eight models:

A 10-second live-action movie trailer for an original video game. The start image is a frame of the game's night street. The reference images show the tea house front, the tea house interior, the three characters, and the two cars.

STYLE: photorealistic live-action film. Turn the street from the start image into a real street at night: wet cobblestones, neon, paper lanterns, strings of red crescent-and-star flags overhead, the same layout and red shop signs. Anamorphic lens, shallow depth of field, light film grain, moody night colors.

CHARACTERS (real actors, identical in every shot):
- Courier: the young man in the middle of the character sheet, plus a dark flat cap as in the start image; navy puffer jacket with a red chest stripe, green cargo trousers, shoulder bag.
- Thugs: black suits and sunglasses, like the man on the right.
- Tea master: the old man on the left: flat cap, grey mustache, brown vest, white shirt.
CARS: the courier's low orange sports car and the thugs' dark grey boxy sedan, exactly as on the car sheet. No emblems, no real car brands.

ONE CRATE: the courier starts with empty hands. He picks up one wooden crate of tea in shot 1 and keeps it until shot 4. It never appears or vanishes by itself.

SHOT 1 (0-3s): The courier walks down the middle of the street, steps to the red stall with the striped umbrella on the left, lifts a wooden crate of tea off its counter with both hands and starts running.
SHOT 2 (3-5s): A dark sedan packed with thugs turns in at the far end, headlights blazing, several thugs leaning out of the windows, and speeds toward him. He glances back, hugging the crate.
SHOT 3 (5-7.5s): Low camera at road level. He puts the crate in the orange sports car parked at the right curb, jumps in and speeds away; the sedan roars past the camera after him.
SHOT 4 (7.5-10s): The orange car brakes in front of the tea house with the big red sign. The courier carries the crate inside and sets it on the wooden counter; the tea master pours black tea into a slim tulip-shaped glass on a saucer, steaming.

AUDIO: driving drums and strings, footsteps, engines, tire screech. No dialogue, no shouting, no voices.
No new text or numbers, no license plates, no logos, no UI.

The skeleton of the prompt: first the style, then a description of each character and car, then the “one crate” rule (the crate must not appear or vanish by itself), then four shots second by second, and finally audio and what’s not allowed.

7. WHAT EACH MODEL MADE

A warning up front: there’s a lot of slop. I’ll say it before you do in the comments.

• Kling 4.0: not bad. There’s consistency between scenes; just a strange turn by the car.
• Seedance 2.5: no logic. We grab the tea, run, get in the car, stop in front of the tea house, but then come out from inside. The thugs look great as always, but I’d have liked them leaning out of the windows; they all stayed inside.
• Grok Imagine 1.5: there’s a physics error at one point, the courier seems to crash into the car. And he gets back into the car that’s already parked in front of the tea house.
• MiniMax H3: the courier guarding the crate and running is good, the thugs are great. But no mustache.
• Gemini Omni 1.1: very realistic, but it has nothing to do with our drawings. There’s no getting-out-of-the-car scene either; I think the problem is my prompt here, I didn’t write that scene.
• Wan 3.0: with these settings it produced a wide format; the others are 16:9, only Wan is 21:9. It’s cheap and it made every scene logical and physically plausible. The crate we pick up at the start is the same crate at the end, and the chasing car is the same car throughout.

In one model the courier flew into the car, jumping in through the windscreen. No physics, but it became my favorite scene.

8. WHICH MODEL WON?

Overall, Kling 4.0, Wan 3.0 and Seedance 2.0 performed well. If I had to continue with a single model to finish my trailer, I’d probably pick Wan, because the connection between scenes is very good.

To be honest, part of the problems come from my prompt. If I’d written it in more detail, we’d probably have gotten better results. But at least I saw the difference in style between the models, and I understood a bit better how to use which video model.

If you want to try the same prompt yourself: try it on Higgsfield. Which one is your winner? Leave it under the video. And if you want more videos on AI architecture or how models work, say so there too.

The full video: https://youtu.be/YeYa0pOtZo4