AI UGC: spend half your time on the starting frame

@maxxmalist, who makes AI ads, puts it bluntly: if you want to get good at AI UGC, spend more than half your creation time perfecting the starting frame. In a follow-up he compares two clips, one from a plain text prompt and one started from a crafted reference image. The text-only clip has more plastic-looking skin, worse lighting, less believable framing and a less consistent structure.

That matches how image-to-video models behave. The first frame fixes identity, wardrobe, light direction, lens feel and composition. The motion prompt can only add to what's already there. If the still looks like stock AI, the video will too.

In practice that means generating a batch of candidate stills, picking the most natural one (real skin texture, believable phone-camera framing, motivated light), fixing hands and text in an image editor, and only then animating it. Keep the motion prompt short and physical: a small head turn, the product raised into frame, a breath before the line.

He also flagged pricing: Wan 3.0 on Higgsfield at about $0.042 per second versus about $0.386 for Seedance 2.5. Those are his figures, quoted during a promotion, so check them against current pricing. If they hold, that gap changes how much you can afford to iterate on drafts.

Guardrail: use fictional characters or people who've given consent, and label synthetic spokespeople where the platform or the law requires it.