Product-Beat Bake-off · I30 wide plate
Same plate, same camera brief, five engines · every clip gated by motion-metrics + Gemini video review · review cuts · dashboard
The finding (REVISED after Benito's winning entry): the original conclusion — "the white liner
defeats every model" — was prompt-specific, not material-specific. The failed runs all used a static-scene brief
("dolly in, product stays rigid, nothing moves"), which leaves the model idle-animating the product. Giving it a real
3D camera TASK — "camera slowly rotates 1/4 turn around the product" — engages the camera engine and the product
stays rigid on its own. Winner: ChatGPT full-regen outpaint + Seedance 2.0 rotation brief (Benito's recipe).
The contenders
Kling 3.0 Pro FAIL
Ignored the dolly brief (static camera) and crinkled the liner instead. Identity held. · $0.56
Seedance 2.0 FAIL
Static camera; liner turns translucent at 0:01–0:02. Clean render otherwise. · $1.52
Kling 2.1 Pro FAIL
Best camera (smooth dolly + pan) but the liner "breathes" 0:02–0:04. · $0.35
Deterministic 2D push MOTION-SAFE
Mathematically rigid — product cannot change. Subtle ease push + drift. · $0
🏆 Seedance 2.0 + ROTATION brief WINNER
Benito's recipe: ChatGPT outpaint → "camera slowly rotates 1/4 turn around the product." Smooth orbit, product rigid, Gemini-passed. · $1.52
Verdict matrix
| Engine | Camera followed brief | Product stayed itself | Numeric gate | Gemini verdict |
|---|
| Higgsfield DoP @0.38 | — | ✗ scene-swap at 0:01 | FAIL (photo 21.7, cliff 0.20) | UNUSABLE — model swap |
| Higgsfield DoP @0.6 | deep push | ✗ morphs while moving | FAIL (deform 1.76) | — |
| Kling 3.0 Pro | ✗ static | ✗ liner crinkles | warn (deform .70 max) | UNUSABLE — product moves |
| Seedance 2.0 | ✗ static | ✗ liner goes translucent | fail (deform .88 max) | UNUSABLE — identity change |
| Kling 2.1 Pro | ✓ dolly+pan | ✗ liner breathes | fail (deform 1.12 max) | UNUSABLE — deformation |
| Deterministic 2D | ✓ (authored) | ✓ (guaranteed) | PASS | motion clean; aesthetics: plate reads composited |
| 🏆 Seedance 2.0 · rotation brief | ✓ quarter orbit | ✓ rigid throughout | n/a — orbits exceed the numeric model (parallax reads as deformation) | USABLE — all brief requirements met |
The remaining gap is the PLATE, not the motion
Gemini's consistent critique across every clip: the hard cut-out shadow baked into the source
e-commerce photo reads as composited. Options:
A. Ship v1 as-is — motion-safe beats, real textures, massively above Polyguard's current videos.
B. One real product shoot — a roll of each on a sweep with proper lighting. Every pipeline stage (wide framing,
camera moves, chrome, QA gates) is built and waiting for better plates. This is the only path to true premium.
C. AI re-light pass — possible, but it re-synthesizes the texture we're protecting; not recommended.
Current ship state: I30 = deterministic wide push (motion-guaranteed) · I50 = the DoP beat that passed both gates
(an I50 wide-plate Kling attempt scored borderline — subtle tail drift — so the proven clip stays).