Engines & quality

The right engine for the job, not the same one for every job.

The AI landscape changes monthly. We test the engines continually, so you do not have to, and we recommend the one that gives the best result for your budget.

Video engines

No single engine is best at everything. Strengths shift between releases, so treat the notes below as our current working view. We will tell you honestly which suits your project.

H3 (MiniMax)

Pros

Strong at holding characters and scenes steady across shots
Native upscaler, so rerolls are cheaper and you upscale only once the take is right
Accepts multiple references and audio, which helps continuity

Cons

Takes careful, disciplined prompting to get the best out of it
Clip length per generation is limited

Seedance

Pros

Good motion and multi-shot cinematic feel
Capable at following detailed direction

Cons

Can cost more per usable second depending on tier
Consistency across many separate shots needs more management

Grok

Pros

Fast and economical for quick drafts and concept tests
Useful for generating reference stills

Cons

Less control for long, continuity-critical sequences
Not our first pick for final TV-level work

Choosing the engine is one of the biggest cost levers in a project, which is why we make that decision with you. See pricing.

Audio engines

Voice is half of a screen performance. We generate dialogue with ElevenLabs, whose expressive voices and direction tags let us shape delivery line by line, and we work with alternatives such as Resemble AI where a project calls for a different voice or licensing approach.

We never artificially speed up or slow down speech to make a line fit. If a line is too long for its shot, we re-roll it or rework the beat, because stretched dialogue sounds wrong.

Character voices cast and kept consistent across episodes
Levels set per delivery type, from whispers to shouts
Music, atmosphere and sound effects built to sit under the dialogue

Quality: 720p to 4K

Higher resolution costs more, and you rarely need it at every stage. We work in the cheapest resolution that lets you judge the shot, then upscale the takes you keep.

Tier	Best for	Relative cost
720p	Drafts, animatics, social clips, testing ideas	Lowest
1080p	Web series, YouTube, most commercial work	Low to moderate
2K	Polished online release, large screens	Moderate
4K	Broadcast, festival and premium delivery	Highest

Why upscaling matters. With engines that include a native upscaler, such as H3, you can reroll at a lower cost until the shot is right and then upscale only the final take. That is usually far cheaper than rendering every attempt at full resolution.

Mature content: mild nudity and violence

Stories sometimes need an edge. Mild nudity and mild violence, the territory of an M (Mature) classification, can be generated. But it is the most demanding kind of work in AI production, and it is worth knowing why before you plan around it.

What the engines allow

Every major engine filters. Sexual content and graphic violence are checked on the way in (your prompt and reference images) and on the way out (the finished clip). Seedance, for example, is currently among the strictest.
Grok has a “Spicy” mode. It is the most permissive of the mainstream engines for mild nudity and intense motion. Its limit is that it can work from Grok’s own generated stills rather than your character and set references, so holding a character steady across shots is harder.
H3 (MiniMax) sits in the middle. Reports suggest it allows suggestive content within limits and blocks anything explicit, and it accepts the reference and start/end-frame inputs we rely on for consistency.
Seedance and Kling are currently the strictest. Both screen the images you supply as well as your prompt, so a reference that trips the filter is rejected before anything is generated.
Open-weight models such as Wan can be run on our own hardware, which is where the in-house generation and upscaling mentioned below comes in.
Rules change often. What works one month can be blocked the next, so we test current behaviour on your project, not last quarter's.

What that means for you

It takes more work. Expect more generations, more rejected attempts and more manual fixing per usable second.
It costs more. Failed generations are often not billed by the engines, but in-house generation and our hours both rise. As always, you pay actual generation costs plus hours.
Some shots need in-house work. Where the commercial engines will not deliver, parts of a shot are generated and upscaled in-house, which adds time.
Consistency is harder to hold when we are working around filters, so we plan for it up front.

Our recommendation: imply it

In most situations we recommend an implied scene. The shot builds right up to that moment, then cuts to a reaction, a detail or the aftermath: a face, a door, a shattered window, the quiet after. It is the oldest trick in cinema because it works. It lets the audience's imagination do the heavy lifting, it keeps the scene inside what the engines handle reliably, and it is usually the most cost-effective option. It is also easier to place on the platforms most viewers watch on.

If your story needs the moment shown, we will tell you honestly what is achievable and what it will cost before any work begins. Everything we produce is mild, M-level content, and we do not produce explicit material or depict real people without their consent.

Tell us what you want to make

A short chat with a real person, in plain English, is the fastest way to find the right engine, quality level and budget.

Start a conversation