Go to content
Menü >>>

Film & Music

Skip menu
Menü >>>
Skip menu

Why generative AI simulates lens physics, yet fails at shadows

Film Txt

The "Cinematic" Fallacy: Why Generative AI Simulates Lens Physics, Yet Fails at Shadows

On the screens of tech evangelists, the future of cinema currently looks like an endless sequence of commercials. Open the relevant channels, and you see breathtakingly rendered faces, neon-lit puddles, and floating tracking shots. Proudly displayed beneath them are the magic formulas of the new era: "shot on 35mm anamorphic, Arri Alexa, 80mm prime lens, natural lens breathing, 180-degree shutter, Kodak Portra 400 grain, masterclass lighting, cinematic photorealism".
The promise is simple: the camera is dead; the text box is the new set.
Yet anyone who has ever stood in a muddy field at dawn trying to balance an HMI light against fading daylight, or spent hours in colour grading separating sensor noise in deep shadows from digital mush, looks at these results with a mixture of fascination and profound disillusionment.

Generative video models such as Sora, Kling, Runway, or Veo produce captivating image sequences. Yet they do not understand cinema. They master the surface, but they fail at the physical and dramaturgical laws of the craft.
  1. The Misunderstanding of the "Cinematic Look"
    The word cinematic has degenerated into the cheapest euphemism of the present day. To software engineers and prompters, it means: a colour palette of teal and orange, an extremely shallow depth of field, and a lighting mood that looks as though the sun has been hanging at a golden angle above the horizon for twelve uninterrupted hours.
    Cinematography, however, is not a colour filter, nor is it a string of pretty postcards. Cinematography is the visual grammar of a narrative. When a director of photography decides where to place the camera, they are not making a choice about what looks good—they are deciding what the audience is permitted to feel and know. More importantly, they are deciding what the audience is not permitted to see. This is precisely where the fundamental design flaw of generative image models begins.
  2. The Horror of the Void: The Absence of "Negative Fill"
    In professional lighting, there is a tool that contains no light source at all: negative fill. Large black duvetyne cloths or floppies positioned on the shadow side of an actor to absorb every unwanted reflection in the room. Cinematographers such as Roger Deakins or Gordon Willis built their life's work on taking light away rather than adding it. Tension in film arises not from what is illuminated, but from the edge where information dies and black begins.
    Diffusion models do not know this concept. They hate information-free black.
    Their mathematical architecture relies on deriving pixel patterns from noise that carry a statistical probability of recognisability. The neural network wants to fill every segment of the frame with structure, texture, fabric folds, and reflections.
    The phenomenon: even in supposedly "bleak noir scenes", a diffuse residual light always glows in the background of AI videos. Every shadow area still exhibits a micro-highlight; every eye has a perfect catchlight; every stretch of tarmac gleams as if after a summer shower.
    The consequence: there is no longer any visual focal point. When every image element competes for the exact same level of attention, the frame loses its dramaturgical hierarchy. The result is not a film frame, but digital baroque—visually overloaded, emotionally sterile.
  3. Optical Physics Versus Pixel Morphing
    To fake cinematic authenticity, the latest model generations deliberately imitate lens imperfections. Prompts now routinely include instructions for lens breathing (the slight shift in field of view when a lens pulls focus from near to far) or anamorphic flares. Yet this is where the difference between physical optics and vector calculation becomes painfully apparent:
    The Circle of Confusion
    A real lens focuses photons through ground glass onto a silicon or celluloid surface. When a point leaves the plane of focus, it dissolves according to the laws of refraction into a blur circle whose characteristics are determined by aperture blades, lens coatings, and optical geometry.
    An AI model does not calculate light rays. It calculates probability vectors of neighbouring pixels. When a character steps forward in an AI clip, the background does not change through optical depth of field, but through a pixel transformation. The background morphs instead of optically defocusing. To the human brain, this generates a subliminal sense of unease—the uncanny valley of focal length.
    The Illusion of Lens Breathing
    When prompters ask AI for lens breathing, the model usually scales the entire image up or down by two per cent. With a genuine Cooke or Panavision lens, however, the image breathes spatially: the foreground shifts its parallax relative to the background differently from the edges of the frame. It is a mechanical movement in space, not a digital zoom.
  4. The 180-Degree Shutter Betrayal
    Every camera trainee learns the 180-degree shutter rule in their first month: at 24 frames per second, each individual frame is exposed for exactly 1/48th of a second. This brief temporal blur during movement is the cadence to which the human eye has been conditioned for over a century. It creates the organic flow that distinguishes cinema film from 1990s camcorder video.
    Generative models do not operate with an exposure time. They generate discrete intermediate states which they smooth out through neural interpolation.
    The result is an unnatural, rubbery fluidity: movements do not smear naturally, but melt like warm wax. Or, during fast camera pans, they appear jerky and harsh because the algorithm tries to hold onto edges that should physically have dissolved into motion blur long ago. The image simulates 24 frames, yet moves with the tactile feel of a 120 Hz video game engine.
  5. The Acoustic Parallel: The Death of Silence in the AI Score
    This misunderstanding of craft is not confined to the image. It continues seamlessly into music and soundscapes. Since tools such as Suno or generative orchestral plugins entered the scene, producers have been flooding their editing timelines with synthetic soundtracks. At first listen, these tracks sound full: beds of strings, heroic brass sections, immaculately mixed sub-bass.
    Yet they commit the exact same error as image generators: they do not understand silence.
    A great film composer is defined not by how many notes they layer beneath dialogue, but by when they allow the orchestra to fall silent. The impact of the opening chord in a Western comes from the dry wind that preceded it for two full minutes. The force of a timpani strike exists only in contrast to the space that swallows it without an echo.
    Generative music engines are trained to stack frequencies without gaps. They deliver musical wallpaper paste designed to concrete over every emotional pause in the edit. Yet anyone who eliminates the space between the notes takes the breath out of the scene.

Conclusion: Craft Is Reduction
Generative video and audio tools are impressive instruments for visual pre-production. For mood boards, colour concepts, and rough layouts, they accelerate processes that once took days. As a replacement for work on set, however, they fail at the exact point where art begins: with restraint.
A cinematographer who knows what they are doing does not light up a room—they sculpt the light until only the skeleton of the scene remains visible. A director does not move the camera simply because they can, but when the scene compels the movement.

A composer holds back the instruments so that the actor's breathing takes the lead. AI can supply us with every conceivable texture. It can conjure up anamorphic lenses that never existed and simulate orchestras that were never tuned.

Back to content
Application icon
Film & Music Website by Barry Redhead Install this application on your home screen for a better experience
Tap Installation button on iOS then "Add to your screen"