Top 3 video models – overall score

    What to use for each task

     

    Best for following prompts and getting the motion you asked for: Seedance 2.0

    Best for camera control and executing specific shots: Seedance 2.0

    Best for physically plausible motion and believable interactions: Flux 3

    Best for videos that stay consistent and clean over time: Flux 3

    Best for human performance and emotional expression: Flux 3 or Grok Imagine Video 1.5

    Best for following an art direction or visual style: Grok Imagine Video 1.5

    Best for using reference images and keeping visuals faithful to them: Grok Imagine Video 1.5

     

    The leaderboard

    Model Video Generation Scores
    Rank
    (overall)    		 Vendor Model Instruction and
    motion adherence
    /10    		 Camera control and
    shot execution
    /10    		 Physical motion and
    interaction plausibility
    /10    		 Temporal coherence
    and visual integrity
    /10    		 Human performance and
    emotional expression
    /10    		 Art-direction
    fidelity
    /10    		 Visual and
    reference fidelity
    /10    		 Overall
    video score

    * model retired

    Our scoring panel is made up of specialist video producers, directors and animators from the Definition video and brand teams.

    They have been creating brand films, motion design and campaigns for brands like Google, DHL, British Gas, FSC and Thomson Reuters, for decades.

    How it works

    We test every major video model across seven core video and production capabilities:

    1. Instruction and motion adherence
    2. Camera control and shot execution
    3. Physical motion and interaction plausibility
    4. Temporal coherence and visual integrity
    5. Human performance and emotional expression
    6. Art-direction fidelity
    7. Visual and reference fidelity

     

    Our specialists, led by senior animator and producer Arshad Peters, review outputs designed to test each capability and grade the model out of 10 for each.

    We add new models to the table soon after their release. Benchmark programme led by Tom Pallot.

    Want access to all of the best video models in one secure place?

    Start a free Definition AI trial today

    More detail on the scores

    When our video team reviews the outputs, they’re assessing each capability against these criteria:

    Instruction and motion adherence

    How well can the model follow a precise brief, with the right objects in the right places, moving the right way at the right time?

    • Subjects, attributes and counts
    • Spatial arrangement
    • Motion accuracy (right object, path, direction and speed)
    • Timing and sequence
    • Exclusions and constraints

    Camera control and shot execution

    How well can the model carry out specific camera moves and focus changes within a single continuous shot?

    • Shot type and angle
    • Locked-off stability
    • Camera move accuracy (correct move, direction, speed and order, with a true dolly rather than a digital zoom)
    • Lens and focus control
    • End framing and continuity

    Physical motion and interaction plausibility

    How convincingly do people, objects and materials move and look the way they would in the real world?

    • Ground contact and weight
    • Gravity and momentum
    • Fluid and particle behaviour
    • Secondary motion
    • Cause and effect

    Temporal coherence and visual integrity

    How well does the model keep people, objects and textures stable and consistent from the first frame to the last?

    • Identity persistence
    • Anatomical integrity
    • Object and edge integrity
    • Texture and surface stability
    • Scene stability

    Human performance and emotional expression

    How convincingly can the model recreate realistic human emotion?

    • Emotional readability
    • Transition and timing
    • Facial nuance
    • Gesture and body language
    • Eyeline and intent

    Art-direction fidelity

    How faithfully can the model deliver a specifically briefed art look?

    • Lighting and time of day
    • Colour grade and colour locking
    • Mood and stylisation
    • Composition, hierarchy and negative space
    • Execution without reinterpretation

    Visual and reference fidelity

    How well (and naturally) can the model use reference images?

    • Product fidelity
    • Logo and text preservation
    • Person identity
    • Environment fidelity and integration
    • Fidelity through motion