MiniMax H3 for Dropshipping Ads: A Test Workflow

Dropshipping is a testing business, and the constraint has always been creative. You cannot film a product you have not received, so testing an angle used to mean ordering samples and waiting. Text to video removes that step: you can test whether an angle sells before anything ships.

This is the workflow that follows from that.

Why does this change product testing?

The old order was: pick a product, order a sample, wait, film it, edit, launch, learn.

The new order is: pick a product, test the angle with generated video, learn, then order samples only for the products where an angle actually worked.

That reversal matters because most products fail on the angle rather than on the execution. You find out cheaply, in an afternoon, rather than expensively, in three weeks. And when a product does work, you already know which promise to build the real creative around.

Step one: find four or five real angles

Open the supplier listing and the competitor ads, and write down why anyone would actually care. Not features, reasons.

For most products you will find four or five: a price angle, a time saved angle, a problem removed angle, a status angle, a novelty angle.

Write each as a sentence a real person would say out loud. "I have thrown out three of these, this one stayed" is a hook. "Premium stainless steel construction" is a spec.

This is the part that decides whether the day is worth anything, and skipping it produces five clips that all say the same thing. Our guide to writing AI UGC ad scripts goes deeper on finding these lines.

Step two: test the angles with no assets

With no reference images attached, MiniMax H3 runs as text to video, so you can generate a person saying each line without owning the product, hiring anyone or booking anything.

Hold everything constant except the line:

A woman in her late twenties sits cross legged on a bed in a sunlit bedroom, holding a small skincare tube toward the camera. She looks straight into the lens and says, "HOOK LINE HERE." Handheld medium close up with a small amount of natural sway, soft window light from the left, clean natural colour. Room tone only, no music. Vertical framing.

Five seconds, 480p, vertical. Generate all five before you judge any of them, then watch them on a phone with the sound off. Keep whatever earns a second of attention.

More templates are in MiniMax H3 UGC prompt templates.

Step three: build the product beats from supplier photos

Once an angle survives, you need shots of the actual product. Attaching one image switches H3 into image to video, where your still becomes the literal first frame, so your real product stays accurate.

Supplier photos work for this when they are clean. Crop to vertical first, because the output follows the image's aspect ratio.

The camera pushes in slowly from a low angle as soft light rises from the left, catching the edge of the product. Shallow depth of field, clean highlights. Quiet room tone, no music.

What does not work: marketplace images with text burned in, collages of several products, or dim images shot on a warehouse floor. Any lettering in the source degrades as the clip moves. If your supplier images are poor, generate a clean product still first and animate that.

The technique is covered in how to use image to video.

Step four: keep one creator across the winning product

When a product proves out, consistency starts to matter. Attaching two to four images moves you into reference to video, where the model carries the person from one image and the product from another through a shot it composes.

Bind them explicitly:

The woman from Image 1 sits at a kitchen table holding the bottle from Image 2, turning it once toward the camera as she says, "HOOK LINE HERE." Medium close up, static camera, soft window light, clean natural colour. Room tone only, no music. Vertical framing.

That gives you a set of variants that look like one campaign rather than ten unrelated videos. Walkthrough in how to use reference images.

Consistency matters within one product's campaign. It does not need to hold across an unrelated catalogue, and forcing one face onto every product is usually worse than matching the person to the buyer.

Step five: finish and launch

Rerun the survivors at 768p, trim the soft frames, add captions, add your price and offer as an end card in the editor. Never ask the model to render text.

Then put two variants live rather than choosing between them, and decide your stopping rule before the first impression.

What does this change about product selection?

It makes killing products cheap, which is the real advantage.

Most of dropshipping is deciding what not to sell. When testing an angle costs an afternoon rather than a sample order and a week, you can run that test across many more candidates, and you stop carrying products through to inventory on the basis of a hunch.

The honest framing: this does not make bad products sell. It makes finding out faster, and it moves the moment of finding out to before you have money committed.

How many products can you test this way?

More than before, and that changes the shape of the business rather than just the creative budget.

The old constraint was that each product test consumed a sample order, a filming session and an edit. That capped most operators at a handful of serious tests a month, which meant product selection had to be right, which meant a lot of deliberation before anything was tried.

When the creative test is an afternoon, you can run it across every candidate on your shortlist rather than the two you liked best. That inverts the deliberation: instead of arguing about which product to try, you try them all cheaply and let the results argue.

A sensible cadence is a batch of candidates per week, five angles each, judged the same day. Most will fail, and failing is the output you paid for.

One caution worth keeping. Cheap testing makes it tempting to test badly, spraying half considered angles at everything. The discipline still applies: five genuinely different promises per product, identical production style, only the line changing. Sloppy tests are cheaper than they used to be and just as uninformative.

What about disclosure and platform rules?

Worth taking seriously. Rules on synthetic media and on ads featuring generated people change regularly and vary by region. Check the current guidance from Meta and TikTok for Business before you run creative featuring a person, and our FTC disclosure guide covers the broader picture.

This is not a reason to avoid AI creative. It is a reason to read the current rules rather than assume last year's still apply.

Frequently Asked Questions

Can I make video ads without having the product in hand?

Yes. Text to video tests the angle with no assets at all, and your supplier photos can be animated as first frames for the product beats.

How many angles should I test per product?

Four or five genuinely different promises. Testing five versions of the same promise with different fonts is not a test, it is a preference survey.

What do I do when a product fails the creative test?

Kill it and move on, which is the point. A cheap creative test is the fastest way to find out a product has no angle before you buy inventory.

Should the creator in my ads be consistent across products?

Not necessarily. Different products often suit different people. Consistency matters within one product's campaign, not across an unrelated catalogue.

Do supplier photos work as reference images?

Often, if they are clean and well lit. Busy marketplace images with text overlays or several products in frame work badly and are worth recreating.

Do I need to disclose that an ad is AI generated?

Platform rules on synthetic media change and vary by region, so check the current guidance from each platform before you run creative featuring a person.

Run it on your next candidate

Take the next product you are considering and test five angles before you order anything. Generation is billed in credits included with your plan, from 1,400 a month upward, with details on the pricing page.