MaxVideoAI Research

AI video model benchmarks, explained

Compare editorial quality scores, sourced model capabilities, and observed generation times in one clear research hub.

Built from MaxVideoAI scorecards, current model specifications, and anonymized production latency.

31

Editorial scores

A consistent 0–10 framework for quality, motion, control, and production fit.

29

Sourced specifications

Current capabilities checked against provider sources and the MaxVideoAI route.

Observed speed

Rolling 30-day median and P90 generation time for eligible models.

Model scorecards

Compare the current MaxVideoAI editorial view across eleven production criteria.

MaxVideoAI editorial score

Compare the current MaxVideoAI editorial view across eleven production criteria.
ModelOverall score (0–10)PromptVisualMotionConsistencyHumanTextAudioSequencingControlStabilityValue
Kling 3.0 Omni ProScore updated · Jun 1, 20268.6
8.7
8.6
8.5
8.5
8.2
6.9
8.8
8.7
9.1
6.4
7.7
Kling 3.0 Omni 4KScore updated · Jun 1, 20268.5
8.6
9.0
8.4
8.5
8.2
7.0
8.6
8.7
9.0
5.6
4.8
Seedance 2.0Score updated · Feb 19, 20268.5
8.6
8.8
8.7
8.2
8.4
7.0
9.2
8.4
8.8
7.4
Happy Horse 1.1Score updated · Jun 22, 20268.3
8.4
8.3
8.3
8.2
8.2
7.0
9.0
7.8
8.3
6.8
8.1
Kling 3 ProScore updated · Apr 25, 20268.3
8.6
8.4
8.4
8.0
8.1
6.8
8.6
8.0
8.7
6.6
7.5
Kling 3 4KScore updated · Apr 25, 20268.2
8.4
8.9
8.2
8.0
8.1
6.9
8.4
8.0
8.5
5.8
4.8
Kling 3.0 Omni StandardScore updated · Jun 1, 20268.2
8.2
8.1
8.2
8.2
7.9
6.7
8.5
8.3
8.8
6.7
8.3
Gemini Omni FlashScore updated · Jul 1, 20268.1
8.7
8.2
8.0
7.7
8.0
7.4
8.9
7.8
9.0
7.1
8.4
Google Veo 3.1Score updated · Apr 3, 20267.9
8.4
8.1
7.9
7.4
8.2
7.2
9.0
7.8
8.3
7.4
7.4
Happy Horse 1.0Score updated · Apr 28, 20267.9
8.0
8.0
8.0
7.8
8.0
6.8
8.8
7.6
8.1
6.8
7.2
Kling 2.6 ProScore updated · Jan 27, 20267.9
8.3
7.9
8.1
7.3
7.7
6.5
8.2
6.5
8.0
6.8
9.2
Kling 3 StandardScore updated · Apr 25, 20267.9
8.1
7.9
8.1
7.6
7.7
6.6
8.2
7.5
8.3
6.9
8.0
Kling 2.5 Turbo7.8
8.1
7.6
8.2
7.1
7.3
6.0
6.0
7.0
7.2
9.6
Seedance 2.0 FastScore updated · Apr 3, 20267.8
8.0
7.8
8.1
7.2
7.8
6.8
8.8
7.6
8.1
8.8
8.8
Luma Ray 3.2Score updated · Jun 11, 20267.7
7.9
8.0
7.9
7.4
7.6
6.4
7.2
8.4
6.6
7.8
Google Veo 3.1 FastScore updated · Apr 3, 20267.6
8.1
7.1
7.7
7.0
7.6
6.5
8.4
7.5
7.9
9.1
9.0
Seedance 1.5 ProScore updated · Apr 6, 20267.5
7.8
8.0
7.6
7.2
7.4
6.8
8.1
7.1
7.5
7.2
8.5
Dreamina Seedance 2.0 MiniScore updated · Jun 16, 20267.4
7.6
7.4
7.6
7.0
7.3
6.4
6.8
7.1
7.7
7.8
9.4
OpenAI Sora 2 Pro7.4
7.6
8.4
8.0
6.7
8.6
6.7
8.3
6.5
7.1
6.2
6.2
Google Veo 3.1 LiteScore updated · Apr 3, 20267.2
7.8
6.7
7.2
6.7
7.3
6.1
8.1
7.3
7.5
9.0
9.8
OpenAI Sora 2Score updated · Jan 27, 20267.2
7.4
8.2
7.8
6.5
8.4
6.5
8.0
6.0
6.8
6.5
9.3
LTX 2.3 ProScore updated · Mar 6, 20267.1
7.6
7.4
7.5
6.2
7.3
6.5
7.8
6.5
7.9
7.8
8.8
LTX 2.3 FastScore updated · Mar 6, 20266.7
7.1
6.9
7.2
5.9
7.0
6.2
7.3
6.0
7.1
9.0
9.2
Luma Ray 2Score updated · Apr 2, 20266.7
7.1
8.3
6.9
6.1
7.4
5.8
5.8
7.8
6.8
6.4
Wan 2.6 Text & Image to VideoScore updated · Jan 27, 20266.2
6.2
5.2
6.2
6.2
5.8
4.8
4.0
5.8
6.5
7.5
9.0
LTX Video 2.0 ProScore updated · Jan 27, 20266.0
6.7
6.0
6.6
4.8
6.8
5.8
7.5
5.5
6.0
7.5
8.9
Luma Ray 2 FlashScore updated · Apr 2, 20266.0
6.4
7.5
6.3
5.4
7.0
5.6
5.4
7.2
8.5
8.6
LTX Video 2.0 Fast5.6
6.2
5.7
6.2
4.5
6.3
5.5
7.0
5.0
5.7
9.2
9.3
MiniMax Hailuo 02 StandardScore updated · Jan 27, 20265.4
5.8
4.8
5.3
5.2
4.8
4.0
4.5
5.8
8.5
9.8
Pika 2.2 Text & Image to VideoScore updated · Jan 27, 20265.4
5.1
5.3
5.6
5.4
5.8
5.0
4.0
6.0
7.5
9.6
Wan 2.5 Text & Image to Video4.8
4.7
4.7
4.9
4.8
5.2
4.5
4.0
5.0
5.5
8.0
9.3

Verified model specifications

Review the limits and workflows exposed through MaxVideoAI.

Review the limits and workflows exposed through MaxVideoAI.
ModelMax durationMax resolutionReleaseInput modesAudioReferencesSource
Dreamina Seedance 2.0 Mini4-15s480p / 720pJun 2026
Text → videoImage → videoVideo → video
SupportedRef2V: up to 9 image references; I2V: 1 start imageSource
Gemini Omni Flash10s720pGoogle preview: Jun 30 2026
Text → videoImage → videoVideo → video
SupportedSupported (up to 10 reference images)Source
Google Veo 3.18s4KOct 2025
Text → videoImage → videoVideo → video
SupportedImage-to-Video: 1 start image; Reference-to-Video: 1-3 stillsSource
Google Veo 3.1 Fast8s4KOct 2025
Text → videoImage → videoVideo → video
SupportedImage-to-Video: 1 start image; Reference mode: 1-3 stillsSource
Google Veo 3.1 Lite8s1080pMaxVideoAI availability: Apr 2026
Text → videoImage → videoVideo → video
SupportedStart image only; no multi-reference stillsSource
Happy Horse 1.015s output (3-60s source for video edit)1080p
Text → videoImage → videoVideo → video
SupportedSupported (1-9 reference stills)Source
Happy Horse 1.115s output1080p
Text → videoImage → video
SupportedSupported (1-9 reference stills)Source
Kling 2.5 Turbo10s1080pSep 2025
Text → videoImage → video
Not supportedSupportedSource
Kling 2.6 Pro10s1080pDec 2025
Text → videoImage → video
SupportedSupportedSource
Kling 3 4K15s4K
Text → videoImage → video
SupportedImage-to-video: 1 source image; optional end frameSource
Kling 3 Pro15s1080p
Text → videoImage → video
SupportedImage-to-video: 1 source image; optional end frame; Kling Elements in promptSource
Kling 3 Standard15s1080p
Text → videoImage → video
SupportedImage-to-video: 1 source image; optional end frame; Kling Elements in promptSource
Kling 3.0 Omni 4K15s4K
Text → videoImage → video
SupportedReference-to-video: @Image references plus Kling Elements; I2V: one start imageSource
Kling 3.0 Omni Pro15s1080p
Text → videoImage → videoVideo → video
SupportedReference-to-video and V2V: @Image references plus Kling Elements; I2V: one start imageSource
Kling 3.0 Omni Standard15s1080p
Text → videoImage → videoVideo → video
SupportedReference-to-video and V2V: @Image references plus Kling Elements; I2V: one start imageSource
LTX 2.3 Fast20s (>10s requires 1080p / 25 fps)4K (6–10s); 1080p for >10s routingData pending
Text → videoImage → video
SupportedSupported (single start image; no separate style-reference stack)Source
LTX 2.3 ProGenerate 6–10s; Audio/Extend/Retake up to 20s4K on T2V/I2V generate; workflow-specific limits for Audio/Extend/RetakeData pending
Text → videoImage → videoVideo → video
SupportedSupported (single start image; no separate style-reference stack)Source
LTX Video 2.0 Fast20s route; standard generate presets are 6-10s4K on 6-10s generate modes; workflow-specific limits for Audio/Extend/RetakeOct 2025
Text → videoImage → video
SupportedSupportedSource
LTX Video 2.0 Pro10s4KOct 2025
Text → videoImage → video
SupportedSupportedSource
Luma Ray 3.210s generation; source clips up to 30s for Modify/Reframe intake1080p
Text → videoImage → videoVideo → video
Not supportedSupported through guide frames and indexed Modify keyframesSource
MiniMax Hailuo 02 Standard10s768pJun 2025
Text → videoImage → video
Not supportedSupportedSource
OpenAI Sora 212s720p (1280×720 landscape; 720×1280 portrait)Sep 2025
Text → videoImage → video
SupportedSupported (single start image; no style-reference stack)Source
OpenAI Sora 2 Pro4 / 8 / 12s720p / 1080pSep 2025
Text → videoImage → video
SupportedSupported (single start image)Source
Pika 2.2 Text & Image to Video10s1080pFeb 2025
Text → video
Not supportedSingle still only; no reference-style stackSource
Seedance 1.5 Pro12sT2V: 480p/720p; I2V: 480p/720p/1080p
Text → videoImage → video
SupportedSingle start image; optional end frame in I2VSource
Seedance 2.015s4K on the standard Dreamina Seedance 2.0 route
Text → videoImage → videoVideo → video
SupportedRef2V: up to 9 image references; I2V: 1 start imageSource
Seedance 2.0 Fast15s720p
Text → videoImage → videoVideo → video
SupportedRef2V: up to 9 image references; I2V: 1 start imageSource
Wan 2.5 Text & Image to Video10s1080pSep 2025
Text → videoImage → video
Not supportedSupportedSource
Wan 2.6 Text & Image to VideoUp to 15s (per generation)Up to 1080pDec 2025
Text → videoImage → video
Not supportedSupportedSource

Observed generation times

Median and P90 end-to-end generation time over a rolling 30-day window.

Kling 3 Pro
Median
3m 51s
P90
7m 32s

Rolling 30 days

How the scores work

One scoring language, one overall formula, and one canonical prompt pack for future documented runs.

Overall score formula

(Prompt adherence + Motion realism + Temporal consistency) ÷ 3

Evaluation scale

2/10

Major visible failures prevent practical use.

5/10

Usable in selected shots with clear limitations.

8/10

Strong production-ready behavior for the tested criterion.

10/10

Exceptional behavior with no material issue in the evaluated outputs.

Prompt adherence

How closely the output follows requested subjects, actions, composition, and constraints.

Visual quality

Perceived detail, realism, lighting, material rendering, and absence of visible artifacts.

Motion realism

Smoothness, physical plausibility, camera movement, and temporal behavior of moving subjects.

Temporal consistency

Identity, object, background, and style stability across the generated clip.

Human fidelity

Faces, hands, limbs, body proportions, contact, and human-object interactions.

Text legibility

Accuracy and stability of requested words, labels, interface elements, and signage.

Audio and lip sync

Alignment between visible speech, dialogue timing, and generated audio when supported.

Multi-shot sequencing

Continuity, narrative order, and identity preservation across explicit shot changes.

Controllability

Reliability of camera, reference, framing, and other supported production controls.

Speed and stability

Editorial assessment of delivery consistency; observed median and P90 latency are published separately.

Value score

Relative production value at the current MaxVideoAI price position; live pricing remains on pricing and generation surfaces.

Canonical prompt pack

01Hands and object interaction

Canonical prompt — English

Documentary-style medium close-up of a ceramic artist at a sunlit studio table. The artist picks up a small blue cup with the left hand, rotates it slowly with both hands, checks the rim, and places it beside a wooden carving tool. Keep five natural fingers on each hand, believable contact, stable cup geometry, and one continuous camera take.

02Two-character identity continuity

Canonical prompt — English

Two distinct friends meet beside a train-station clock at golden hour. One wears a red denim jacket and round glasses; the other wears a navy wool coat and carries a folded paper map. They exchange the map, point toward platform seven, then walk together toward the camera. Preserve each face, outfit, body position, and the map throughout one continuous shot.

03Fast motion and camera tracking

Canonical prompt — English

Low-angle tracking shot beside a cyclist moving quickly through a wet city street just after rain. The camera accelerates smoothly, passes a reflective bus shelter, then arcs around the cyclist without changing the bicycle frame, wheel shape, clothing, or direction of travel. Preserve realistic tire contact, water spray, motion blur, reflections, and stable urban geometry.

04Product and text legibility

Canonical prompt — English

Premium studio product shot of a matte black coffee can rotating slowly on a dark stone pedestal. The front label must read exactly “MAX NORTH” in large white letters with “COLD BREW” directly below in smaller gold letters. Use a controlled rim light and a gentle push-in. Keep the can cylindrical, the typography readable, and every letter unchanged for the full shot.

05Cinematic lighting and materials

Canonical prompt — English

Cinematic night-market scene in light rain, photographed at eye level with a slow forward dolly. Warm paper lanterns reflect on wet pavement while steam rises from a stainless-steel food cart. A vendor in a green apron serves a ceramic bowl to a customer. Preserve realistic skin, fabric, metal, ceramic, steam, rain, reflections, and consistent lighting without flicker.

06Constrained camera control

Canonical prompt — English

Perfectly centered architectural shot of a modern concrete gallery with a single red chair in the middle. Begin with a locked symmetrical wide frame for two seconds, then perform one slow straight dolly forward. Do not pan, tilt, roll, orbit, zoom, add objects, move the chair, or change the lighting. Keep vertical lines straight and the composition symmetrical.

07Dialogue and lip synchronization

Canonical prompt — English

Medium close-up of a presenter in a quiet daylight studio looking directly into the camera. The presenter says exactly: “Every frame should feel intentional, clear, and ready to share.” Use a natural conversational pace, clean studio sound, stable facial identity, realistic blinking, and precise lip movement synchronized to every spoken word. No subtitles and no background music.

08Three-shot sequencing

Canonical prompt — English

Create a coherent three-shot sequence featuring the same woman in a yellow raincoat and the same silver compact camera: shot one, a wide view as she enters a foggy pine forest; shot two, a close-up as she raises the camera and takes one photograph; shot three, an over-the-shoulder view of the captured deer on the camera screen. Preserve her identity, clothing, camera, weather, direction, and color grade across all shots.

Method notes

  • Generative outputs vary between runs.
  • Not every model supports every criterion or prompt type.
  • Provider capacity and queues change over time.
  • Observed production traffic is not a controlled experiment.
  • Rolling performance may reflect incidents, routing changes, and user-selected settings.
  • MaxVideoAI sells access to the compared models and discloses that commercial interest.

Methodology updates

  1. v1.0.0

    Initial score definitions, overall formula, canonical prompt pack, and rolling latency rules.

Failed paid generations are automatically refunded.