Phantra AI: Product Shots
Creates product-led Phantra scenes anchored to the user’s product photos, with macro and turntable shots, and explains 4:5 or 1:1 reformat options.
Product shots
A product scene is held to the person's own product photos: the exact
object, label and colour on screen come from them, and the scene's job is to
light it, move around it and show its surfaces. The clip is a scene, written
here free, saved with write_scene, quoted, and rendered on the person's yes.
Get the product in as a prop
elements: is the product there as a prop with photos? Itsdescriptionsays what the studio read off them (name, colour, size).- If not, ask the person for it.
add_elementwithkindprop, anameandsource: the https address they gave you of the product's own page (its main picture is taken, and what the page states about the product is kept in the notes as copied text) or of an image file. A photo on their device comes in throughupload_linkwithelement_kindprop andelement_name: hand them theurl, they pick the file, thenelementsagain. One picture per call; the same name again adds to it. Never giveadd_elementa picture you found yourself. - Two or three photos help: a clean pack shot (front, label readable) and one showing the top or the back. Ask rather than guess what the back looks like.
Roles and anchor
On write_scene the pack shot is the FIRST ref with role product: the
render anchors on it. A photo attached for its lighting alone (a mood frame,
a surface) gets light, palette or style and goes after; a look-only
photo is never the anchor. A room photo gets place.
The style block: surfaces and reflections
Name the surfaces, because they are what the model renders: the product's own finish (matte card, brushed metal, glass with visible thickness, frosted plastic) and what it sits on (wet slate, linen, a mirror, raw concrete). Say where reflections come from and how sharp they are: one large soft source gives a long soft highlight down a bottle; a hard point source gives a small bright one. Say the background (a seamless sweep, a dark falloff, a real counter) and the lens (a macro a few centimetres away, or a 50-85mm at table height with shallow depth). Say "label faces the lens and stays readable". No brand is named for the look; the product's own name is on its label.
Shot patterns (each window ONE shot, 2-10 s)
- Macro drift: the lens a few centimetres from one detail (cap, seam, embossed name), a slow push or slide, focus racking from texture to label.
- Turntable: the product on a slow rotating surface, camera locked, a half or full turn in 4-6 s, the highlight sliding across it as it turns.
- Reveal: a hand sets it down, a lid lifts, mist clears; one action, then still.
- Hero hold: locked three-quarter view, a slow 10-degree orbit, light wrapping the edge.
Two or more windows, contiguous from 0 to the total (4-30 s; 8-12 is usual). Sound is diegetic: "No background music. Only diegetic sound: the surface, the cap, the pour, the turntable's quiet motor."
Worked skeleton (replace everything in it)
Ultra-realistic product film in 9:16. The attached photo of the serum bottle
is the exact product, label and colour. Style: amber glass with a visible
wall and a brushed-metal cap; one large soft source high left giving a long
soft highlight down the bottle, a hard rim from behind right; wet black
slate with a sharp inverted reflection; dark seamless falloff; macro at 5 cm
for details, 85mm at table height for the whole; shallow depth, true colour,
no bloom.
(0-4s) Macro on the cap's brushed grain, a slow slide down to the label,
focus racking onto the name. Label faces the lens.
(4-10s) Hard cut: the bottle on a slow turntable, locked 85mm, a half turn,
the highlight sliding round the glass, the reflection turning with it.
No background music. Only diegetic sound: the turntable's faint motor, one
drop of water on slate.
Avoid: plastic AI sheen, invented text on the label, extra bottles, lens
flare, named brands for the look.
Save, quote, approve, and the shape afterwards
Show the prompt, take edits, then write_scene (prompt, seconds, refs
pack shot first, roles, title, project_id). quote it: keyframes
(stills x each), clip (a render per timed shot, each with a token) and
balance. Say them in the quote's own numbers -- the model, the shots and
their credits, the keyframes, the balance -- and WAIT for an explicit yes; only then approve
(what clip or keyframes, the quote's tokens, its renderer choice), job
until done; scene shows clips. A still is generate_image (the prop's
ref in references; image_models says which take them): the first call is
the quote only, the same call plus its quote_token the draw, once. Never
approve for the person or reuse a token; refused: insufficient_credits means
stop: credits are added in the studio. Once a clip has landed, reformat (its
gen:<id> from library) makes a copy in 4:5 or 1:1 for a feed (free, a
job, the source untouched): fit pad keeps the whole picture with black bars
where the shapes differ; fit crop fills the frame and keeps the middle, which
is why the product should sit near the centre of the scene.
- Downloads
- Security audit
- Pass
- Last updated
- 2w ago
- Current version
- v1.0.0
- License
- MIT-0
