Hailuo AI Video Prompt Generator: Directing With Camera Tags

MiniMax’s Hailuo models are unusually explicit about cinematography: the family is known for responding to named camera techniques, which makes it the best destination for a reference clip you want to imitate, because a reference already contains nothing but named camera techniques. Here is the tag vocabulary and how to extract it from footage.

Why this family is camera-tag driven

Most models let you describe a shot in ordinary language and hope. This one documents and rewards a director’s vocabulary: specific movement names, shot sizes and angles behave like switches. The practical consequence for our workflow is excellent — a reference clip converted into a tag list becomes a prompt that reads as instructions rather than as wishful description. The trade-off is that a vague prompt wastes the model’s strongest capability.

The tag vocabulary

Category	Tags that behave
Shot size	extreme close-up, close-up, medium, wide, establishing
Angle	low angle, high angle, eye level, overhead, dutch tilt
Movement	push in, pull back, pan left/right, tilt up/down, tracking shot, aerial, orbit, crane
Stability	handheld, steady gimbal, locked-off
Pacing	slow motion, time-lapse, single continuous shot
Optics	shallow depth of field, wide-angle distortion, macro, anamorphic flare

One tag per category, and never two movements — that rule matters more here than anywhere, precisely because the model obeys the tags so literally. The physics behind the words is in extracting camera movement with AI.

Extracting tags from a reference

This is the shortest conversion of any model family, because an extractor’s output maps almost one-to-one onto the tag list. Paste the clip URL into SocialToPrompt (27 platforms, or upload MP4/MOV/WebM to skip the extraction credit) and read back the camera, composition, lighting and pacing fields; the run takes the stated 20–40 seconds.

Then convert to tags and verify the movement by eye. On a reference of a coffee commercial shot, that looks like: close-up, low angle, push in, locked-off base with slight drift, shallow depth of field, warm practical light.

What extraction cannot give you, and what you must add: the duration, the end state, and the one thing you want to change relative to the reference.

Assembling the prompt

[subject + action]  [tag list: size, angle, movement, stability, optics]  [light]  [grade]
[duration]  [ending]

A barista presses espresso into a portafilter, steam curling past the rim. Close-up, low angle,
slow push in, slight handheld drift, shallow depth of field. Single hard window key from
camera-left, deep shadow behind. Warm muted grade, fine grain. Six seconds, ends on the
tamper lifted from frame.

Subject and action lead — models weight what comes first — and the tags follow as a compact block rather than being woven into prose. Light and grade are stated, never implied, because a reference-derived prompt that omits them defaults to a generic look.

Hailuo-specific failures

• Two movements, one smear. If the reference has a dolly-in followed by a pan, generate two shots and cut them.
• Over-tagged optics. Wide-angle plus anamorphic flare plus macro is a contradiction; pick the dominant one.
• Tag words in the wrong slot. Movement buried after a long subject clause loses to the subject; keep the tag block early.
• Inherited motion from a stabilised reference. Gimbal footage extracts as static more often than not — check it before accepting an empty movement tag.

For the same reference aimed at other dialects, see Seedance, Kling and Wan; the underlying extraction is what a prompt extractor returns. And the same rights test applies to every model family: camera technique is not owned, expression and likeness are.

Frequently asked questions

How do I write a prompt for a Hailuo-family model from a video?

Extract the clip's fields and convert them into a compact tag block: shot size, angle, one movement, stability, optics, then light and grade. Lead with the subject and action, keep the tags early and literal, and add the duration and ending yourself, because no extractor supplies those.

How many camera tags should one prompt contain?

One per category and, critically, only one movement. This family follows camera instructions unusually literally, so two movements in a short clip produce an averaged smear rather than a sequence.

Why is my generated shot static when the reference wasn't?

Because stabilised gimbal movement often extracts as no movement — the background shifts so smoothly that the motion reading reports a locked camera. Watch the frame edges on the reference and write the movement tag by hand.

Do mood words help with this model family?

Minimal value. Named techniques and physical light descriptions do the work; adjectives about atmosphere are the least-weighted content in the prompt. Spend your words on optics, direction of light, and what changes.