What you are watching
This is an unedited example of local text-to-video output — the kind of quick motion test you would run before committing to a longer generation or refining the prompt. It was rendered on the creator's own GPU rather than submitted to a hosted video service, which is the entire point of the tool that made it.
It is deliberately short. Diffusion video models are expensive to run, and a three-second clip at a modest resolution is the sensible unit for iterating on a prompt: long enough to judge motion and coherence, short enough that you are not waiting on a slow render to find out the prompt was wrong.
Clip details
Resolution
Native output resolution, not upscaled after the fact.
Frame rate
Typical for short diffusion video clips, where every frame is generated.
Duration
Roughly three seconds — a standard length for a motion test.
Codec
MP4/H.264, so it plays anywhere without a plugin.
Generated on local hardware
Saient runs video generation models on your own machine. You describe the shot — subject, motion, lighting, camera behaviour — choose the frame count and step count, and the model denoises the clip locally. Nothing about the prompt or the output is uploaded.
The same approach runs on a phone. Saient's Android app generates Wan2.1 text-to-video clips end-to-end on the device's own Vulkan GPU, which is unusual enough that we wrote it up in detail.