AI ads are polarizing, but the data doesn't lie. Major ecom brands are pushing 90% of spend through AI creatives now because the testing velocity is absurd - 10x faster than traditional shoots. Cheaper, quicker, but here's the trap: most of what floods feeds is absolute slop. Run bad AI and you're torching brand trust faster than a clickbait thumbnail.

My process shifts every few weeks as models update, but this is what's converting right now. Long post, but I'm laying out the exact pipeline, not some growth hack.

Reference images are everything
Text-to-video is a lottery. First-frame control on Kling or a solid reference image on the others determines whether the generation holds up or falls apart at second three. I pull model + backdrop combos from Pinterest that match the brand vibe, then run them through Nanobanana to swap the face and background. Critical: the Pinterest shot is purely compositional reference. We don't want real people. NB outputs look way cleaner when it has a visual anchor. Every prompt specifies "iPhone camera, natural lighting, raw texture" - generations are at the point where you genuinely can't tell they're synthetic. I repeat this for every single scene. Reference quality = generation quality.

Model selection: three worth running

Omni - newest, cheapest, fastest, surprisingly consistent. My workhorse for UGC talking heads. Lip sync holds, voice stays stable, follows reference frames faithfully. Does sometimes leave that subtle AI sheen on female avatars, but a correction layer in post fixes it. Wouldn't touch it for complex motion, but for static delivery? Unbeatable efficiency. (VEO still feels like a beta to me.)

Kling - love-hate relationship. Frustrating inconsistency, but for B-roll? It's the king. Product-in-hand shots, avatar demos, close-ups - Kling nails the physics. Mid-tier cost and speed. I only spin it up for specific B-roll slots.

Seedance - quality ceiling is highest, no debate. Also brutally expensive and slow. Used to run everything through it until Omni matured. Now it's a specialist tool: hero shots only. Burning 20 minutes and heavy tokens on a single 5-second clip that might glitch? Not sustainable for volume testing.

Prompting is where retention lives or dies
Perfect reference + weak prompt = wasted render. I trained a custom GPT that's processed hundreds of generations - it thinks in structure now. Here's the exact framework it outputs:

Quality/Fidelity Lock - identity preservation, smartphone realism, zero beautification, no HDR, no cinematic grading, no influencer polish. Question: what should this feel like?
Reference/Identity Lock - exact person, clothing, environment elements. Question: who is this and what's frozen?
Scene Lock - location, props, lighting, time of day. Question: where are we?
Camera Behaviour - this is where 80% of generations fail. Camera type, framing, movement style, operator intent. Locked talking head (zero drift), UGC selfie (natural arm motion), observer (friend filming), participant (inside the moment). Question: who's holding the phone?
Performance/Action - body movement, eyeline, energy, gestures, intent. Question: what's happening?
Dialogue - exact words, delivery style, lip sync cues. Question: what do they say?
Audio - biggest recent unlock. Phone mic realism, dominant sound, what shouldn't be heard. Question: what does the mic pick up?
Style Intent - memory, ad, story, documentary, TikTok, camera roll. This stops the model from outputting a commercial. Question: what format is this?
Negative Constraints - every recurring model error. No captions, overlays, beauty filters, hero shots, cinematic movement, staged behavior. Question: what must never appear?

Formula runs in that exact order. Character limits vary by model so I sometimes cap the GPT's output. Always feed the reference image into the chat alongside the prompt - consistency lives or dies there.

Post is where AI becomes an ad
Pacing and texture. Built a correction LUT that strips the AI glaze - night-and-difference difference. Dialogue speed drifts per generation, I speed up almost everything to keep hook retention past the 3-second cliff. No single clip holds screen longer than 2-3 seconds. Swap angles, cut to B-roll, change the avatar's environment 2-3 times per creative. Captions must be frame-accurate and styled for silent autoplay.

Bottom line: you can master every model on this list and still fail if the ad doesn't work. Pretty hook, zero relevance = immediate swipe. The generation skill is table stakes. The advertising skill is what scales.

Questions welcome - happy to go deeper on any stage.