Most of the "AI look" comes from three defaults models fall into when you do not override them: even sourceless light, frictionless camera movement, and a clean digital surface. Override all three and the same model produces footage people stop questioning. None of this requires a better model, and all of it is cheaper than switching.
1. Name the light source and direction
The highest-leverage change available in a prompt. Models default to pleasant ambient illumination that comes from nowhere, and nothing in the real world is lit that way.
Instead of: "cinematic lighting, beautiful, high quality" Prompt: "single hard source from camera left, late afternoon, deep unfilled shadow on the right of the face, warm 3200K"
Specify the source, its direction, its hardness, its colour temperature, and what is left dark. The last one matters most: committing to shadow is what makes a frame look lit rather than rendered.
2. Prompt the imperfection
Generated camera moves are too smooth, so ask for the flaw.
Useful phrases: handheld with slight drift, small overshoot at the end of the pan, minor focus hunt before settling, camera settles after the push, slight vertical bob from walking.
For UGC-style work push further: phone camera, mixed indoor lighting, slightly off-centre framing, autofocus correction mid-shot.
3. Break the eye contact
Sustained unblinking eye contact with the lens is the strongest uncanny trigger in the format. Prompt the glance away and back, the blink, the look down. Or frame three-quarter rather than straight on.
4. Grade it, always
Raw model output has a flat, slightly plastic look. Even a light grade moves it a long way: set a black point, warm or cool it deliberately, and let something clip. Footage where nothing is fully black and nothing is fully blown reads as synthetic because real capture rarely behaves that politely.
5. Add grain
Small amount, applied after grading. Digital cleanliness is a tell and grain is the single cheapest correction available. Match the grain across a sequence or the mismatch becomes its own tell.
6. Mix in one real shot
The most underused technique. One piece of real footage in a sequence raises the perceived reality of everything around it, because the viewer calibrates on the real frame and extends the benefit of the doubt.
A phone clip of the actual product, one real hand, one real establishing shot. Cheap, fast, and worth more than any prompt change.
7. Cut it shorter than feels right
Generated clips reveal themselves over time: artefacts accumulate, motion drifts, the shot starts announcing itself around the four-second mark. Trim to the two seconds that work and cut against something else. Most generated footage improves dramatically by being used less.
A worked prompt
Weak: "A woman drinking coffee in a kitchen, cinematic, 4k, beautiful lighting"
Strong: "Medium shot, woman lifting a ceramic mug, kitchen. Single window source camera left, overcast midday, soft but directional, right side of her face falls into shadow, no fill. Handheld with slight drift, small settle at the end. She glances down at the mug, then out of frame left. Shallow depth, 50mm feel. 5 seconds."
The second names light, direction, shadow, camera imperfection, eye behaviour, lens and duration. It is longer because each clause removes a default.
What this cannot fix
Physical causality, hands doing precise work, and identity across a long take. Those need a real shot. See why AI video still looks fake for where the line sits.
FAQ
Does a longer prompt always help? A more specific one does. Length that adds adjectives does not; length that removes defaults does.
Should I use negative prompts? Where supported, yes, for recurring artefacts. Naming what you do want works better than listing what you do not.
Is grain cheating? It is finishing, which is what every piece of real footage also gets.
Name the light, ask for the flaw, grade it, and cut it short. The 8frame canvas is free and unlimited, and generation is paid from $19/month.