18 AI video models, one shot: what each model did
We gave 18 AI video models the same start image and the same 10-second camera move. See what each model did, beat by beat, with every clip and its price.
Asaf Naim ·
We gave 18 AI video models the same start image and the same prompt: one 10-second camera move through an impossible crystal city. On this take, only MiniMax H3 hit every beat of the move and kept the world of the start image. No model cut, but the models split on everything else: where the camera went, whether the woman stayed still, and what the city turned into.
- Models
- 18
- Clips
- 19
- Clip length
- 10 s
- Start frame
- 1
- Beats checked
- 6
- Hard cuts
- 0
The test
Every model got the same start image, the same prompt and a clip length of 10 seconds. Everything else stayed at each model's default, including the resolution and the sound. Each model made one take. Grok Imagine (I2V) made two takes, and this page shows both. We ran the test in September 2026.

The start frame comes from our image comparison, where 15 image models drew the same scene. This is the prompt every video model received, word for word:
We checked every clip against six beats from the prompt, in order:
- The camera pushes toward the woman and moves past her.
- The woman stays standing at the edge.
- The camera flies into the city, between the glass towers.
- The camera passes beneath the floating islands.
- A whale passes beneath the camera.
- The camera dives under an island and reveals a figure walking upside down.
We also checked whether the clip keeps the world of the start image: the night, the red moon and the crystal city.
The start frame has no upside-down figure in it. A model that shows one at the end had to invent it.
What each model did
✓ means the clip does it, ~ means it does part of it, and ✗ means it does not. The models are in alphabetical order, not ranked.
| Model | 1 | 2 | 3 | 4 | 5 | 6 | World | Credits |
|---|---|---|---|---|---|---|---|---|
| Bytedance Lite I2V | ✓ | ✓ | ✓ | ✓ | ✓ | ~ | ✓ | 63 |
| Bytedance Pro Fast I2V | ✗ | ✗ | ✗ | ✗ | ~ | ✗ | ~ | 99 |
| Bytedance Pro I2V | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ | 193 |
| Gemini Omni | ✓ | ✓ | ✓ | ~ | ✓ | ✗ | ✗ | 495 |
| Gemini Omni Flash | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | 352 |
| Grok Imagine (I2V), take A | ~ | ✗ | ✗ | ~ | ✓ | ✓ | ✓ | 88 |
| Grok Imagine (I2V), take B | ✓ | ✓ | ✗ | ~ | ✓ | ~ | ✓ | 88 |
| Grok Imagine Video 1.5 | ✓ | ✓ | ✗ | ✗ | ✓ | ✓ | ~ | 407 |
| Hailuo 2.3 Pro | ✓ | ✓ | ✗ | ~ | ✓ | ✗ | ~ | 220 |
| Hailuo 2.3 Standard | ✓ | ✓ | ~ | ✓ | ✓ | ~ | ~ | 143 |
| Kling 2.5 Turbo Pro | ✓ | ✗ | ✓ | ✓ | ✓ | ✗ | ✓ | 231 |
| Kling 2.6 | ✓ | ✗ | ✓ | ✓ | ~ | ✗ | ✗ | 303 |
| MiniMax H3 | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | 1,012 |
| Runway Gen-3 | ✓ | ✓ | ✗ | ~ | ✓ | ✗ | ✓ | 33 |
| Seedance 2 | ✓ | ✓ | ✓ | ✓ | ✓ | ~ | ✓ | 627 |
| Seedance 2 Fast | ✓ | ✓ | ✓ | ✓ | ✓ | ~ | ✓ | 517 |
| Seedance 2 Mini | ~ | ✓ | ✗ | ~ | ✓ | ✗ | ✓ | 319 |
| Seedance 2.5 | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ | 770 |
| Wan 3.0 Prime | ✓ | ✓ | ✗ | ✓ | ✓ | ✗ | ✗ | 693 |
Credits is today's price of the same run: a 10-second clip at the resolution the model used. Seedance 2, Seedance 2 Fast and Seedance 2 Mini charge in steps of 4, 8, 12 and 15 seconds, so a 10-second clip costs the 12-second price. Each model's page lists all of its prices.
The clips came out at each model's default resolution, from 736×400 to 2560×1440. Compare what each model did, not how sharp the picture is.
What we saw
No hard cuts, but two scene swaps.
Scene-cut detection found no hard cut in any of the 19 clips. Gemini Omni and Gemini Omni Flash still swapped the crystal city for a modern glass city, in a dissolve between two and four seconds in.
10 of 19 runs flew into the city.
MiniMax H3, Seedance 2, 2 Fast and 2.5, Bytedance Lite and Pro, both Kling models and both Gemini models flew between the towers. Eight runs never entered the city.
Four runs moved the woman who should stand still.
Kling 2.6 walked her down the walkway, and Kling 2.5 Turbo Pro flew her off the platform. Grok Imagine take A floated her down among the whales, and Bytedance Pro Fast walked her toward them.
Four runs invented a figure walking upside down.
MiniMax H3, Gemini Omni Flash, Grok Imagine Video 1.5 and Grok Imagine take A ended on it. Five more ended on a small figure the right way up: on top of an island, under a structure, or hanging from one.
Four runs left the world of the start image.
Both Gemini models built a modern glass city. Wan 3.0 Prime flew out of the crystal to a castle city above the clouds, and Kling 2.6 turned the night into bright day.
12 of 19 clips came with sound.
The prompt says nothing about sound. Both Gemini models, the Grok models, Kling 2.5 Turbo Pro, MiniMax H3, the Seedance models and Wan 3.0 Prime added sound by default.
The whale was the easiest beat: in 17 of the 19 clips, a whale passes beneath the camera. The hardest part of the prompt was its last line, the dive under an island to a figure walking upside down.
Every clip
Each clip below is the model's original file, at its own resolution. A clip loads only when you press play. The picture before you play is the clip's last frame. The six frames under each clip show the camera's path, from 1.5 to 8.5 seconds, left to right and top to bottom.
Alibaba

Passes her, then leaves the crystal for a castle city above the clouds. The red moon shrinks to a small disc.
ByteDance

Pushes past her into the city and under the islands, with a whale close below. It ends on a small figure standing on top of an island, not under it.

The camera barely moves. The woman walks toward the whales, and the glass beside her breaks into a kaleidoscope pattern.

A smooth, steady flight past her and into the city. It ends on a floating island against the moon, without the dive.

Follows the path into the city and under the islands, with a whale close below. It ends under a huge structure, where a tiny figure stands upright.

Takes the same path as Seedance 2. It ends inside a structure lit by beams of light, where a tiny figure stands upright.

The view changes early to ring-shaped platforms. The camera drifts among them with the whales and never enters the city.

A slow, steady flight past her and through the city, with whales below. It ends under an overhang, without the figure.

The crystal city dissolves into modern glass skyscrapers about three seconds in. The camera passes her, and whales swim between the towers to the end.

Dissolves into a modern glass city about two seconds in, and still hits every beat. It ends under an island, on a figure walking upside down.
Kuaishou

A very fast flight, with streaks of light. The woman flies off the platform, and the clip ends on the red moon and the whales.

The woman walks down the walkway. The camera flies into the city, the night turns into bright day, and the clip ends on a whale up close.
MiniMax

Passes her and drifts into mist among the whales. It ends beside a floating island.

Brighter and hazier than the start frame. It ends under an island, where a small figure hangs from it, head up.

Pushes past her, flies through the city and under the islands, and dives beneath one. It ends on a figure walking upside down beneath the island.
Runway

Passes her and drifts toward the moon and the whales. It ends on a whale beside a floating island.
xAI

The woman floats off the platform and down among the whales. It ends under a sheet of glass and falling water, on a figure walking upside down.

Passes her and skims over the whales, outside the city. It ends under a cracked glass canopy, where a figure hangs from it, head up.

Passes her, then turns to the moon and the whales. It ends on a dark figure upside down against a clean red moon.
How we read the clips
We checked every clip frame by frame: at least 8 frames spread across each clip, and 24 frames across the few seconds where a beat was unclear. A beat counts as done when the clip shows it clearly, and as partly done when the clip shows a version of it. For example, a figure under an island that hangs the right way up is partly done.
Scene-cut detection measured how much each frame changes from the one before it. No clip had a change large enough to be a hard cut. A dissolve changes the picture gradually, so we found the two Gemini dissolves by watching.
This comparison is one take per model, on one start frame and one prompt. A second take can turn out differently, as the two Grok Imagine takes show. Read the table as what each model did with this shot, not as a ranking of the models.
Run the same test in Nodaro
You can run a comparison like this one on your own shot. Connect one image and one Text node to several Generate Video nodes. Choose a different model in each Generate Video node, set the same clip length, and run the workflow. Each node shows its price on its Run button before it runs. To compare image models first, do the same with Generate Image nodes.
Frequently asked questions
Related
15 AI image models, one prompt: what each model drew
Generate Video
Video models
Camera Motion Lab: how 147 videos taught us to move a camera
Choosing a model
Last updated on
Compare
Side-by-side tests of AI video and image models in Nodaro. Every model gets the same prompt and settings, and every result is shown with today's price.
15 AI image models, one prompt: what each model drew
We gave 15 AI image models one long prompt with 15 hard details. See which details each model drew, from the crystal city to the figure walking upside down.