Vivenso Labs

Research — AI video models

Four video models, put to two tests.

First, make a clip from a written brief. Second, take a picture that already exists and set it moving as instructed. Both tests were run on an architectural scene, an action scene and a scene from a fictional setting. Every clip is below, side by side, as each model returned it. Press play on any of them.

Run on 2 October 2026 One run per model Movement not judged

The short version

What each model did.

JobBest in this runWhat we saw
Make a clip from a detailed briefGemini Omni Flash, then Veo 3.1Omni showed every listed item in all three scenes; Veo missed the side view of the runner
Move a given picture as instructedSeedance 2.0, when it accepts the pictureIt made the camera move and the spear thrust it was asked for, and declined one picture of three
Knowledge of a fictional settingVeo 3.1 and Gemini Omni FlashBoth drew the two warriors as described; the other two put a helmet on the bare-headed one
SpeedGemini Omni FlashUnder a minute a clip; Seedance 2.0 took about five
Price per clipKling 3.0The cheapest clip, and the only one returned at 1920 by 1080

Whether the movement looks right is not judged here. A table can say what is in a clip. Only watching it says how it moves, and that is what the clips are for.

How far to trust it

One brief, one run, checked by frames.

Each model made each clip once. A second run could come out differently, and with video the spread between runs is wide. The tables were filled in by reading still frames taken from each clip, which shows what appears and roughly when. It does not show pace, weight or smoothness.

The models

Four models from three makers.

ModelMakerClip lengthSize returnedTime to makePrice per clip
Gemini Omni FlashGoogle10 seconds, fixed1280 by 720About 46 seconds$1.03
Veo 3.1Google6 seconds1280 by 720About 1 minute$2.40
Kling 3.0Kuaishou5 seconds1920 by 1080About 2 minutes$0.70
Seedance 2.0ByteDance5 seconds1280 by 720About 5 minutesAbout $1.55

Every model received exactly the same words. Each was asked for a wide picture and the shortest clip it offers of five seconds or more; Gemini Omni Flash has no length setting. Every clip came back with sound made by the model, which nobody asked for. The Seedance price is our estimate from its published rate. The clips on this page are the originals, re-encoded at 1280 by 720 so they load quickly.

Test 1

Making a clip from a brief.

Each brief lists things that must appear and how the camera should move. The tables record which of them the frames show.

Test 1 · Architectural

A house on a cliff, approached from the air.

The brief: a slow aerial shot moving toward a two-storey house of concrete and cedar slats, cantilevered over a coastal cliff at dusk, with a reflecting pool on the terrace, waves below, and every line of the building staying straight.

Gemini Omni Flash · 10 s
Veo 3.1 · 6 s
Kling 3.0 · 5 s
Seedance 2.0 · 5 s
Asked forGemini Omni FlashVeo 3.1Kling 3.0Seedance 2.0
Two storeys, concrete and timberYesYesYes, the timber as a thin screenYes
Reflecting pool on the terraceYesYesYesYes
On a cliff above breaking wavesYesYesYesYes
Camera moves toward the houseYesYes, circling as wellBarely movesYes

Test 1 · Action

A runner, tracked from the side.

The brief: a woman running along a coastal trail at sunrise, the camera beside her at hip height and moving at her speed; navy long-sleeved top, black shorts, a watch on her left wrist; no text and no logos.

Gemini Omni Flash · 10 s
Veo 3.1 · 6 s
Kling 3.0 · 5 s
Seedance 2.0 · 5 s
Asked forGemini Omni FlashVeo 3.1Kling 3.0Seedance 2.0
Seen from the side, camera moving with herYesPartly, it swings round to face herYesYes
On a coastal trailYesYesNo, a road with a guard railYes
Navy top and black shortsYesHard to tell, she is lit from behindYesYes
A watch on her wristYes, and she glances at itYesCannot be seenYes
No logosYesYesYesNo, marks on the top and the shorts

Test 1 · A fictional setting

A duel from a world the models were not built for.

The brief named two kinds of warrior from the Warhammer 40,000 setting: a gold-armoured guardian, bare-headed, with a spear, and a red-armoured warrior in a horned helmet with a motorised axe. The spear was to parry the axe in a burst of sparks, at real speed. None of the four refused.

Gemini Omni Flash · 10 s
Veo 3.1 · 6 s
Kling 3.0 · 5 s
Seedance 2.0 · 5 s
Asked forGemini Omni FlashVeo 3.1Kling 3.0Seedance 2.0
Gold guardian, bare-headed, with a cloakYesYesNo, he wears a helmetNo, he wears a helmet
Red warrior in a horned helmet, with an axeYesYesYesYes
Spear meets axe in a burst of sparksYesYes, a large burstNo sparks in the framesYes
A live-action lookYesYesCloser to a game cinematicCloser to animation
Shots in the clipSeveral, with cutsOneOneTwo

On this set

The clips of this setting, here and in the second test, are unofficial test clips made by AI models. Warhammer 40,000 is a trademark of Games Workshop Limited. Vivenso Labs is not affiliated with Games Workshop, and nothing here is endorsed by it.

Test 2

Setting a given picture in motion.

Every model was handed the same starting picture and the same instruction. A good result does what the instruction says and keeps the picture's subject as it was. The starting picture is shown first in each set.

Test 2 · Architectural

Rise, and move in over the pool.

The instruction: the camera rises slowly and moves forward over the pool toward the house, while the water ripples and the waves move. The house itself must not change.

A concrete and cedar house on a coastal cliff at dusk, seen across a reflecting pool.
The starting picture every model was given.
Gemini Omni Flash · 10 s
Veo 3.1 · 6 s
Kling 3.0 · 5 s
Seedance 2.0 · 5 s
Asked forGemini Omni FlashVeo 3.1Kling 3.0Seedance 2.0
Camera rises and moves toward the houseBarely movesYesBarely movesYes
The house stays the sameYesYesYesYes

Test 2 · Action

Keep her running.

The instruction: she keeps running at the same fast pace with the camera tracking beside her, and her face, hair, clothes, shoes and watch stay the same.

A woman in a navy top and black shorts running on a coastal trail at sunrise, seen from the side.
The starting picture every model was given. It was made by an AI model; she is not a real person.
Gemini Omni Flash · 10 s
Veo 3.1 · 6 s
Kling 3.0 · 5 s

Seedance 2.0 declined this starting picture, saying it may show a real person. It does not: the picture was made by an AI model. No clip was made.

Seedance 2.0 · no clip
Asked forGemini Omni FlashVeo 3.1Kling 3.0Seedance 2.0
She keeps runningYesYesYesDeclined the picture
Camera stays beside herYesDrifts round toward her frontYes, pulling backNo clip
Same person and clothes throughoutYesYesYesNo clip

Test 2 · A fictional setting

One fast thrust, then back to guard.

The instruction: he drives the spear forward in one fast, precise thrust at full speed, then snaps back to his guard, with no slow motion, and his face and armour stay the same.

A bare-headed warrior in gold armour and a crimson cloak holding a long spear level, in a ruined hall lit by fires.
The starting picture every model was given.
Gemini Omni Flash · 10 s
Veo 3.1 · 6 s
Kling 3.0 · 5 s
Seedance 2.0 · 5 s
Asked forGemini Omni FlashVeo 3.1Kling 3.0Seedance 2.0
One thrust, then back to guardYes, after holding still for over a second; then he stands for the restNo. He lifts the spear overhead and turns awayA thrust, then a spin that was not asked forYes, at once
Same face and armourYesYesYesYes

Three of the four held the starting pose for more than a second before anything moved. Whether any of these thrusts is fast and precise enough is a question for the eye, not for a table.

Method

How the test was run.

  • The same words for everyone. Nothing was tuned for any one model, and nothing was re-run to get a better clip. There are twenty-three clips: twenty-four were asked for and one was declined.
  • Starting pictures. The house and the guardian are the source pictures of the image study, cropped to the wide shape. The runner was made new in that shape by a model from a maker not under test.
  • What was declined. Seedance 2.0 declined the runner's starting picture as a possible likeness of a real person. The instruction was not reworded to get round that. It accepted the guardian's picture, which also shows a realistic face.
  • One interruption. The two Google models stopped part-way through on a spending-rate limit on our account. That is not a refusal. Their four remaining clips were made twenty minutes later with the same words.