A reel that sells a product has to do three things in under thirty seconds: show the product clearly, show a believable person reacting to it, and give the viewer a reason to stop scrolling in the first second. Most teams can manage one of those with the budget they have. Shooting a product cleanly is a half-day. Hiring a presenter is a week of scheduling and a rate card. Cutting it together is another afternoon.
The workflow below collapses that into an afternoon total, using three tools in sequence: product photography first, presenter second, motion last. The order matters more than the tools do. Generate in the wrong order and you spend your day fixing continuity errors instead of making content.
Lock the Shot List Before You Generate Anything
A 30-second vertical reel is six to nine clips. Each clip runs four to six seconds because that's the useful length most video models produce before motion starts drifting. If you don't decide those clips in advance, you will generate forty seconds of pretty footage that doesn't cut together.
Write the list as plain sentences on one line each. A working structure for a product reel:
- Hook (0–1.5s): the presenter's face plus a visible problem or reaction. No product yet.
- Reveal (1.5–5s): product enters frame, held, label legible.
- Detail (5–10s): macro of texture, pour, click, closure — whatever the product physically does.
- Use (10–20s): presenter actually using it, two angles.
- Proof (20–26s): result, before/after, or a second person.
- Close (26–30s): product on a clean surface, space for an end card.
Decide your aspect ratio now too: 1080 × 1920, and keep anything important out of the top 250 pixels and the bottom 420 pixels, because that's where the platform UI and captions sit. Generating a beautiful centred composition and then discovering the price tag is under the Follow button is a half-hour you don't get back.
Shoot the Product First and Treat It as the Fixed Reference
Everything downstream inherits from the product image. If the bottle has a slightly different cap in clip three, no amount of good editing hides it. So make the product plate first, in Product Photo Studio, and make it the one asset you never regenerate.
Upload the cleanest photo you own — phone camera on a white sheet is fine, as long as the label is sharp and the lighting isn't yellow. Cut it out with background removal before you do anything else, so the model composites onto a clean alpha rather than hallucinating edges from an old shadow.
Then generate the set you'll actually need, not a gallery:
- One hero on a neutral surface, three-quarter angle, soft key from camera left.
- One in-context shot (bathroom shelf, kitchen counter, desk) matching the scene your presenter will be in.
- One macro at the texture or mechanism.
- One held-in-hand plate if you can get it.
A prompt that produces a usable in-context plate:
Product on a warm oak bathroom shelf, soft morning window light from the left, shallow depth of field, matte ceramic and linen props slightly out of focus behind, neutral beige palette, 85mm look, label fully legible and facing camera.
Check the label every single time at 100%. Image models smear small type and invent extra letters. If the lettering is wrong, fix it with the natural-language editor by masking the label area and re-specifying it, or composite your original label back in. A reel with a misspelled brand name on screen is worse than no reel.
If the source photo is small or soft, run it through resolution upscaling before the video stage. Video models downsample, then the platform compresses again — you want headroom.
Cast a Presenter You Can Use Fifty More Times
The value of an AI presenter isn't the first video. It's the fortieth, when the same face has become something your audience recognises. That only works if you treat casting as a one-time decision and build a reference sheet.
In AI Influencer Studio, define the character in specifics that survive regeneration: approximate age, hair length and colour, skin tone, face shape, a distinguishing feature or two, and a wardrobe palette. Vague descriptions drift. Specific ones hold.
Woman, early 30s, warm olive skin, shoulder-length dark brown hair with a centre part, round gold stud earrings, light freckling across the nose, relaxed cream linen shirt, natural makeup, friendly and understated.
Generate a character sheet before you generate any content: front, three-quarter left, three-quarter right, and one full-body. Save them. Those four images are your continuity reference for every clip, every future reel, and every static post. If the tool supports a seed or character ID, write it down somewhere you'll find it in six months.
Match the presenter to the product's buyer, not to a generic idea of attractive. A skincare reel presented by someone whose skin shows no texture at all reads as an ad instantly, and the engagement drops with it. Ask for pores, a little asymmetry, imperfect lighting. The guide to building an AI influencer that actually sells goes deeper on persona and voice if you're setting one up to run long-term.
One practical note: check the disclosure rules for your platform and region before you publish. Several jurisdictions now require you to label synthetic human presenters in commercial content. A small on-screen label costs nothing and protects the account.
Put the Product in the Presenter's Hands Before You Animate
This is the step most people skip, and it's the one that saves the most time. Do not ask a video model to generate a person and an accurate product in motion. You will get a bottle that morphs, a label that melts, and a hand with a surprise sixth finger.
Instead, build still frames where the product and the presenter already exist together correctly, then animate those frames. Composite in the image editor or with a Kontext-style instruction:
Place the bottle from the reference image into her right hand, held at chest height, label facing camera, match the scene's soft left-side lighting and add a contact shadow on the fingers.
Build one still per clip in your shot list. Six to nine stills. Check each one for the usual failure points before moving on:
- Fingers wrapping the product correctly, with the thumb in a plausible place
- A contact shadow where hand meets object — without it, the product floats
- Consistent light direction across every frame in the same scene
- Label sharpness and spelling
- The presenter's face matching the character sheet
If the outfit needs to change between scenes, outfit try-on is faster than re-describing the whole character and risking facial drift. Change the clothes, keep the face.
Animate Each Still With One Instruction Per Clip
Now open the video generator and feed it your stills as first frames. Image-to-video gives you control that text-to-video can't: the composition, the product and the face are already locked, and the model only has to supply motion.
The rule that fixes most bad output: one motion per clip. The camera moves or the subject moves, not both, and not three things at once. Models handle a single intention well and compound instructions badly.
```
First frame: presenter holding bottle at chest height.
Motion: she lifts the bottle slightly and turns it a few degrees toward camera, small natural smile. Camera static. 5 seconds.
```
```
First frame: macro of cream on fingertip.
Motion: slow push-in on the fingertip, cream catches the light. No subject movement. 4 seconds.
```
Use real camera vocabulary — push in, pull back, slow pan left, static locked-off, handheld drift — because those terms appear in training data and map to actual results. "Dynamic cinematic movement" maps to nothing.
Model choice matters for the clips that carry human motion. Faster models are fine for product-only shots like a pour or a rotating hero. For anything with a face talking or hands manipulating an object, step up to something like Kling 3.0, which holds human anatomy together longer. If you're not sure which to commit credits to, the comparison of Wan 3, Flux 3 and Kling 3 breaks down where each one earns its cost.
Budget for a 3:1 generation ratio. You will throw away roughly two clips for every one you keep, usually because of a hand, a blink that lands wrong, or a label that softened mid-shot. That's normal, and it's still faster than a shoot.
Cut for the First Second, Then Add Sound
Assemble in whatever editor you already use. The edit decisions that change performance are few and blunt.
Start on the frame with the most information. Not a logo, not a slow fade — the presenter's face already mid-reaction, or the product already in hand. If your first frame could be anyone's video, you've lost the scroll.
Cut on motion. When a clip ends with her hand moving right, start the next one with movement already in progress. The cut disappears. Hold shots between 1.5 and 3 seconds in the first half, longer in the second half once attention is earned.
For audio, generate a voiceover or a bed track with the audio generator and mix the voice around −6 dB with music at −18 dB. Then burn in captions anyway, because most of your audience watches muted. Two type families maximum, high contrast, positioned above the bottom 420-pixel safe zone.
One colour pass across the whole timeline at the end. Clips generated in separate sessions drift in white balance, and a single LUT or a matched contrast curve across everything is what makes six separate generations read as one shoot.
Turn the Winner Into a Template
Once a reel performs, the expensive part is already done — the character sheet, the product plates, the shot structure, the prompt wording that worked. Save those as a folder and a text file, and the next reel is a swap job rather than a build.
Keep a running prompt log with the exact strings that produced keepers, the seed or character ID, and the model you used for each clip type. Six weeks later, when you need a variant for a new SKU, that file is worth more than any of the finished videos. Change the product plate, keep the presenter, re-run the same six motion prompts, and you have a second reel before lunch.
And before you publish any of it: watch the whole thing once on a phone, at arm's length, with the sound off. Every flaw that survives that test is a flaw your audience will see.
