Depth-of-field prompts that change an AI video frame

Name a focal length and what is sharp. "Cinematic bokeh" is a vibe; "85mm, product in focus, background two stops down" is a frame. Pair it with the model that actually holds the still — Seedance 2.5 I2V via the image-to-video tool, not a restage.

This page is one job: write a depth-of-field clause the model can execute, then animate the picture that already has that plane. A text-to-video restage redraws the bottle and the blur. Focal-length words change geometry. This is the sharpness half of the same sentence.

"Cinematic bokeh" is not a lens

Depth of field is a plane, not a grade. "Cinematic," "dreamy," and "creamy bokeh" tell the model to smear something. Sometimes it smears the product. Sometimes it smears the subject's ears and leaves the wall sharp. The clip looks expensive until you pause on the hero.

A working clause names four facts: focal length in millimetres (85mm, 24mm, macro — not Shot on RED); what is sharp (the perfume bottle); what is soft (the kitchen behind it); how far down (background two stops down or f/1.8). Aperture notation does not control a real iris. Write it with the plane, not instead of it.

Weak: cinematic bokeh, dreamy shallow depth of field. Strong: 85mm, camera 80cm from the bottle, product in focus, background two stops down, f/1.8, the window falls soft.

The second prompt is a frame. Lighting prompts belong in the same cinematography slot: source, direction, quality. Keep the block at the front, next to shot size and angle. Late camera language is the first thing a model approximates away.

The clause that actually changes the frame

Write one still, then change only the lens line. That is the test of whether the words are doing geometric work.

Same subject, same room:

24mm wide-angle, camera in the doorway, deep focus, the product on the table and the window at the back of the room are both sharp

85mm portrait, camera two metres back, product in focus, background two stops down, the window is a soft shape

If both stills show the same hallway, the lens token is decoration on that model. If the wide version stretches the room and the long version collapses it, you have a prompt you can reuse.

Do not stack contradictions. 85mm, everything in focus, cinematic bokeh asks for portrait compression and landscape depth at once. Macro is a thin plane — if you need the whole pack sharp, you want 50mm, f/8, not macro.

If background type is garbage — a shop sign that says almost-a-word — do not inpaint the sign. Soften the band it lives in. A real 85mm at f/1.8 would not have rendered that sign readable. A moving plane is a different job: depth of field and rack focus. This page stays on a held plane.

Hold the still, then add motion

Text-to-video invents the subject and moves it. That is two jobs. When the blur is the brief, the first job is a still you can pause. Then spend the video credit on motion only.

The live model URL is Seedance 2.5, not a folded image-to-video slug. Catalog facts: 4–30 seconds, 720p, 47 credits, native audio. Image-to-video is the tool: AI image-to-video treats the upload as frame one. Seedance 2.5 is the family you pick on that tool so the still holds.

The motion prompt describes camera and physics, not a second product:

Slow 10-degree orbit, hold focus on the bottle, background stays two stops down, no cut, no restage

Do not re-describe the label or the room. Re-describing is how the model draws a cousin bottle with a different falloff. If the take goes deep-focus by second eight, shorten the clip. Drift of a shallow plane is a length problem.

Lock the still before you spend on motion is the routing rule. Seedance 2.5 will run to 30 seconds if you ask; start at 6–10. The AI video generator is the text-to-video door when you want a restage. A plane you already approved is not that brief.

A three-take check

Run this before you build a prompt library on an assumption.

Take one — still, 24mm. Same product, same table, 24mm wide-angle, deep focus, product and window both sharp.

Take two — still, 85mm. Identical prompt except the lens line: 85mm, camera two metres back, product in focus, background two stops down. Compare the window, not the bottle. The wide take should show mullions. The long take should show a soft colour field.

Take three — image-to-video on take two. Upload the 85mm still to image-to-video, pick Seedance 2.5, motion-only, 6–10 seconds. If the background sharpens by the last second, cut duration; do not add adjectives. If the bottle morphs, you restaged.

Pass: the two stills disagree on geometry, and the clip agrees with the 85mm still. Fail on take two: switch the still model. Fail on take three: you used a restage door. Reuse one focal length per scene.

FAQ

Does the word "bokeh" do anything by itself?

As a texture nudge, yes — it biases toward out-of-focus highlights. It does not name a plane. Use it after 85mm, product in focus, background two stops down, never instead of that clause.

Why Seedance 2.5 instead of generating the clip from text?

Because the still is the contract. Text-to-video invents a new bottle and a new blur. Seedance 2.5 through image-to-video keeps frame one: 4–30 seconds, 720p, 47 credits, native audio. Length is optional. The hold is not.

Can I rack focus in the same prompt?

Not reliably as a camera event. A held plane is a state. A rack is a change of state. Stage it: subject walks into the plane, first-and-last-frame, or two clips dissolved. The rack-focus how-to is that page.

What if the background stays sharp no matter what I write?

The lens token is not doing geometric work, or the prompt contradicts itself (85mm plus everything sharp). Run the 24mm vs 85mm still test with nothing else changed. If both files match, stop spending video credits on a clause the still row cannot see.