Shot on an 85mm has become a magic phrase in AI video prompts — pasted at the end of everything, understood by almost nobody. Here’s the awkward truth from my own testing: the phrase works, sometimes. But not for the reason people think. And on some shots it actively works against you.
A lens is not a filter or a vibe. Focal length decides geometry: how big faces render, how near and far layers relate, how much room the frame can hold from where the camera stands. Depth of field decides emphasis: what’s sharp, and therefore what the viewer is told to look at. When you name a lens in a prompt, you’re asking the model to reproduce that geometry — and models, trained on oceans of captioned photographs where “85mm” sits beside portraits, often can. This guide is what the numbers actually do, when to name them, and how to get the same results when naming them fails.
Lenses Are Characterisation
Every lens choice is a statement about the subject before the subject does anything. Wide lenses put the viewer inside the subject’s space — close, involved, the world stretching around them. Long lenses watch from a distance, compress the world behind the subject, and flatter. Which is its own kind of statement. Cinematographers cast lenses the way they cast actors: the lens decides whether we stand with a character or observe them, whether a room feels generous or claustrophobic, whether a face reads as honest or heroic.
That’s why “85mm look” became shorthand for “good portrait.” But shorthand without understanding misfires. I’ve seen creators prompt “85mm” on a cramped kitchen scene, then fight the model for three generations because the frame couldn’t fit the room. No 85mm on earth fits that room from inside it. The model knew; the prompt didn’t. On our storyboards, lens gets written beside shot size (“MCU, 50mm feel, shallow”) as one line of the shot’s specification — because the two decisions are married. Shot size says how much of the subject; lens says from how far and with what spatial flavour. Our storyboard-first workflow keeps them on the same line for exactly that reason.
A quick calibration note. Focal lengths below assume the full-frame reference every photography caption uses. You don’t need the physics. You need three behaviours — wide stretches, normal tells the truth, telephoto compresses — plus one rule about faces. The next three sections take them in order.
Wide Lenses: Energy And Distortion
Wide lenses (roughly 16–28mm) see more from closer. Space opens up: a small room becomes navigable, a street gains sweep, a POV gains hands and world at once. Movement feels bigger too — walk toward a wide lens and you grow fast, which is why wides carry energy, chase, comedy, dance. For AI generation, wides have a second virtue: they’re honest about environments, and models render rooms and landscapes with real conviction at wide framings.
The cost is distortion, and it lands on faces first. Perspective isn’t actually created by the lens — it’s created by camera distance — but wide lenses invite you in close, and at close distance the nearest features (nose, forehead) render disproportionately large while ears and background recede. That’s the “wide-angle face” viewers recognise from front-camera selfies — and in a dramatic scene it reads as clownish unless clownish is the job. Working rules: keep faces near the centre of a wide frame, where distortion is mildest. Never park a face at a wide frame’s edge, where stretching is severe. And use close wide faces deliberately — comedy, menace, POV honesty — rather than by accident.
Prompt shapes: “wide-angle lens, inside the small workshop, camera close, space opening behind him” for energy and place; “18mm feel, low angle, skateboard moving hard toward camera” when exaggeration is the point. If a generation returns faces warped where you wanted straight drama, the fix is rarely a different adjective. Move the prompt’s implied distance back: add “camera a few steps back” or switch the lens family entirely.
The Honest Middle: 35mm And 50mm Thinking
The middle focal lengths earn their reputation quietly. A 35mm shows a person and their immediate world in natural proportion — the documentary and street-photography standard. Wide enough for context, tight enough for attention. A 50mm renders space close to how an eye takes it in: no stretch, no squeeze, faces honest at conversational distance. If in doubt, this family is the doubt-remover. Most narrative AI video that “looks like a film” is, geometrically, 35–50mm thinking wearing other clothes.
For creators, the practical uses split cleanly. Reach for 35mm language (“35mm lens, subject and environment in balance”) when the place matters as much as the person: the cook in her kitchen, the mechanic in his bay, the stallholder in his market. The frame holds a subject at medium distance plus a working margin of world on every side — the classic storytelling frame, and the safest default for two-shots and walk-and-talks. Reach for 50mm (“50mm lens, natural perspective, gentle background softness”) for interviews, dialogue coverage, any face that must be believed: proportions stay true, the background begins to soften without melting, and the viewer feels seated across the table rather than floating in the room.
The middle family’s AI virtue is reliability. Both lengths dominate film and photography training data, so models reproduce their geometry with the least fuss and the fewest warped hands at the margins. When a sequence keeps coming back “almost right” on other lenses, re-boarding it in the honest middle frequently ends the fight. No shame in that. There’s a reason working cinematographers live between 35 and 50 for most of a career. Flattery is a choice, not a default.
Telephoto: Compression And Flattery
Telephoto lenses (85mm and beyond) narrow the view and — the famous effect — compress space: background layers stack up close behind the subject. A street behind a face feels dense and near. A crowd feels pressed in. Mountains loom over a village they’re actually far from. Faces flatter under telephoto for the honest reason that the camera stands farther away: from three metres, nose and ears are nearly the same distance from the lens, proportions even out, features sit in their most regular arrangement. Add the telephoto family’s natural shallow depth of field and you get the portrait look the internet pastes everywhere: subject crisp, world melted to colour.
Use it where flattery and isolation are the job: beauty, romance, product heroes, a face the story wants us to fall for, a single figure the crowd must not touch. “85mm lens, subject isolated, background compressed and soft” is a complete instruction, and models honour it well for faces at close-shot distances.
Now the trade-offs, honestly stated. Telephoto flattens, and flat is wrong when the scene needs spatial energy — a dance, a workshop, a chase through a market will feel embalmed at 85mm. Compression also lies about distance, so avoid telephoto when the geography must be believed (a character “far from the exit” reads close at 135mm). And long-lens prompts on wide shots confuse models legitimately, because the combination barely exists in real footage. If your wide establishing shot keeps returning mushy, check whether a stray “85mm” in your style tail is sabotaging it. Take the magic phrase out of the tail and put lens words only where the shot earns them. I keep them in the camera block of the prompt, beside shot size — the position of honour described in the anatomy of a great AI video prompt.
Depth Of Field: Focus As A Highlighter
Depth of field is the zone of acceptable sharpness, and directing it is directing attention. Shallow depth of field (fast lens, wide aperture — the f/1.8 end of the world) says “look here, only here”: the subject glows out of a soft field, distractions dissolve, intimacy rises. Deep focus says “everything matters”: the foreground clue and the background face are both evidence — as in a mystery or a busy family kitchen. Neither is more cinematic. They’re different sentences.
In generation, focus language is among the most reliable controls you have, because focus is visually unambiguous in training data. “Shallow depth of field, background melting to soft colour” and “deep focus, foreground to background sharp” both land consistently across Seedance 2.5 and Veo. Focus also does composition work that crops can’t: it separates depth planes (the flatness cure from our depth-layers method), it hides background detail the model would otherwise have to maintain over time — a soft background can’t morph noticeably, which is a free stability win — and it flatters skin by letting texture stay natural rather than etched.
The discipline is consistency and dosage. Within one scene, keep a focus logic: if shot one isolates faces in cream-soft shallows, shot three shouldn’t suddenly render every jar on the shelf unless the camera moved somewhere new. Across an edit, mismatched depth is a subtle continuity break viewers feel in their shoulders. And resist maximum shallowness as a default: at extreme blur, eye-lines get hard to read, two-shots lose their second person, products lose the context that sells them. Choose the shallowest focus that still lets the shot say what it must. Focus is a highlighter. A page highlighted entirely is a page not highlighted at all.
Naming Lens And Aperture In Prompts Without The Gimmicks
So: should you write “85mm, f/1.8” in prompts? Sometimes. With eyes open about what the words do. Focal-length words work because captions pair them with consistent geometry; aperture words work because captions pair them with consistent blur. Both are labels for looks the model has seen — requests for a rendering mode, not settings on a virtual camera. Nothing about “f/1.8” changes exposure in a generator; it selects the shallow-focus family of images. Knowing that keeps you from stacking labels superstitiously.
My working rules, after a lot of wasted generations:
- Name the effect first, the number second. “Natural perspective, gentle background softness, 50mm feel” survives a model that ignores numbers; the reverse does not.
- One lens per shot, in the camera block. Not in the style tail, not repeated three ways. Contradictory lens words (“wide-angle 85mm”) produce averaged geometry — the mush again.
- Skip aperture numbers unless blur is the point. “Shallow depth of field” says it in human words; keep “f/1.8” for portrait work where you want the model deep in that mode.
- Have a fallback description for every number. Wide = “space stretches, camera close”. Telephoto = “compressed background, subject isolated”. If a model shrugs at “35mm”, the description carries the shot anyway.
- Never let lens words fight shot size. “Extreme wide shot, 200mm” argues with itself. Size and lens agree first; style words come later, if at all.
A last honest note: lens language can’t fix a badly boarded shot, and it won’t rescue a weak subject description. It’s the final ten percent — the difference between a face rendered and a face characterised. Get size and action right, place the shot in a layered frame, then let the lens cast its vote. Do that, and the numbers stop being gimmicks. They become what they are on a real set: the quietest, strongest directing decision in the prompt. For how these choices tie into finishing, our camera movement prompts for AI video guide completes the camera picture — movement, energy, and now lens, all speaking the same language.