Synthesys

Most AI video tools wrap a single model in an interface. Synthesys does something different: it treats the models as interchangeable parts and picks one per scene.

That distinction matters because video models are not equally good at everything. One holds photorealistic faces, another handles cinematic camera motion, another renders text without turning it to soup.

The rest of the product is a set of format collections designed to constrain the prompt box, which is usually where marketing video generation goes wrong.

Synthesys: One Agent, Six Video Formats

• Multi-model orchestration — the agent selects the model per scene rather than forcing everything through one.
• Six format collections: UGC, hybrid VSL, product and brand, animated, reel ads and commercials.
• Founded in London in 2020, claiming over 1,000,000 videos generated for 50,000+ creators and businesses.
• 1,000+ avatars, 400+ voices, 140+ languages with lip-sync, and avatar videos up to 30 minutes.
• Ratings displayed openly: Trustpilot 4.4, Capterra 4.6, G2 4.7, Software Advice 4.6 and a Product Hunt Product of the Day.

Synthesys: Formats And Ratings

What Does Orchestrating Frontier Models Mean?

The phrase sounds like marketing until you understand what it replaces. A single-model tool sends every request to the same generator, so its weaknesses become your weaknesses on every video you make.

Synthesys states that it selects the optimal model for each scene — photorealistic avatars from one, cinematic motion from another — and composites the result.

That structure has two real consequences. The first is that quality varies less predictably by task, because a weakness in one model can be routed around rather than worked around.

The second is upstream improvement. When Sora, VEO, Kling or Wan ships a better version, an orchestration layer inherits it without rebuilding anything. That is a defensible position in a field where model quality changes every few months.

The cost is reproducibility. If routing is automatic, the model that produced a result you liked may not be the one that handles the next request, and neither you nor the platform necessarily tells you which was used — the same trade every multi-model platform makes, including ImagineArt.

An orchestration layer inherits every improvement the frontier models ship. It also inherits their retirements.

Six Formats And Which Ones Generate Well

The six collections are not cosmetic. Each carries assumptions that do work a prompt would otherwise have to describe.

UGC and reel ads are the strongest fit for generation, because the format's visual language tolerates imperfection. Handheld framing, uneven lighting and imperfect continuity read as authentic rather than broken.

Avatar-led formats — hybrid VSL, spokesperson, training — work well because the hard part is a talking head holding a face across a long take, which is now solved technology rather than an open research question.

Product and brand work is the honest weak spot for any generative tool. A generated object that is nearly your product is worse than useless in an ad, which is why product-first tools like Picwand exist as a separate category from general video generators.

Cinematic commercials are the highest-variance category. When they work they are remarkable; when they fail they fail in the uncanny middle, and the difference is often invisible until you watch on a large screen.

Reading A Ninety-Five Percent Cost Claim

Synthesys states that it reduces video production costs by 90 to 95 percent and delivers output ten times faster, comparing against agency and freelance production it prices between one thousand and twenty thousand dollars per video.

The arithmetic is not wrong. It is the baseline that needs examining.

A 90 percent saving against an agency is only a saving if an agency was your alternative. For a marketing team that has never commissioned a shoot, the honest comparison is against not making the video at all, and the saving is zero because the spend is new.

The claim also excludes the cost that generation adds: reviewing output, discarding failures, and the time spent iterating toward something usable. Every generative tool carries that tax and none of them price it.

Where the claim is most defensible is variation. Producing fifty versions of an ad for testing is genuinely near-impossible at agency rates and genuinely cheap here, and that is a real change in what a team can do rather than a discount on what it already did.

• Compare against what you would actually have spent, not agency rates.
• Budget for review and discard time, which generation does not remove.
• Use the format collections rather than the open prompt box.
• Test variation volume — that is where the economics genuinely change.
• Check commercial rights before anything runs as paid media.

Reading The Displayed Ratings And Logos

Synthesys publishes its scores across four review platforms rather than one, which is a small honesty signal worth crediting.

Trustpilot 4.4, Capterra 4.6, G2 4.7 and Software Advice 4.6 is a consistent picture across audiences that review for different reasons, and consistency is more informative than any single high number.

The site also displays enterprise logos including Coca-Cola, TCS, Yahoo, Heat and Control, AT&S and Jetex, plus press placements. What a displayed logo means is rarely specified — a single team running a pilot and a company-wide deployment look identical on a logo wall.

The percentage outcomes attributed to customer segments are attributed to platform data from 2025 rather than to independent study, which the site states directly rather than obscuring.

None of this is a criticism. It is the difference between what a vendor page can establish and what only your own trial can, and the ratings are strong enough that the trial is worth running. Consistency across a whole body of work is a separate matter again, and closer to what Hypernatural is built for.

Disclosure And Synthetic Presenters

With 1,000-plus avatars, digital twin creation and 140-plus languages with lip-sync, the practical question stops being whether you can make the video and becomes whether you should say how you made it.

The site does not address disclosure, and no vendor page in this category does, so the decision falls entirely to you.

Three things make it worth deciding deliberately rather than by default. Advertising regulators in several markets have begun treating undisclosed synthetic presenters as a consumer information issue. Platforms increasingly require synthetic media labelling on paid placements. And audiences notice — often not consciously, but in the drop-off.

There is a second consent question specific to digital twins. A cloned presenter is a real person's face and voice, and the release you need covers not just the recording session but every future use of the likeness, including scripts written after they leave.

None of this argues against generating the video. It argues for settling two things before the first campaign: what you will disclose, and what your talent has actually agreed to.

Producing Marketing Video Without A Studio

The failure mode with a tool this broad is generating fifty things and shipping none. Six collections invite an afternoon of sampling.

This sequence starts from one real brief and tests the specific claims that decide whether the platform pays for itself.

1. Pick one real brief with a defined audience and placement, not a test prompt.
2. Choose the collection that matches the format rather than starting from scratch.
3. Feed it your actual input — a product URL or image beats a written description.
4. Generate three variations and judge them on a phone, where they will be watched.
5. Time the full loop including review and discards, then compare to your real baseline.
6. Confirm commercial rights and disclosure requirements before the ad runs.