Voxta docs

Video Generation

Services that turn a prompt or a picture into a short clip, locally or hosted, plus the media processing services that upscale, cut out and smooth images and videos.

Video Generation services make a short clip from a prompt (text-to-video), from a picture (image-to-video), or both. Some local models also render a soundtrack with the picture.

Where you use it

  • Create › Visual › Video: the Video playground. Clips are made in projects: New project, then each project has a Generate view and a Gallery view. Pick the service, the duration, resolution and aspect ratio. Add pictures (Frames, meaning start and end frames, or References, when the model takes them), and generate. A draft renders a rough preview first. Render full makes the clip from the same prompt. Continue on a clip picks up where it ends (with some models, motion and sound carry across the join). Join continuations adds a continued clip to the one it came from. Improve my prompt has an AI rewrite the prompt before generating.
  • The image viewer: Animate into a video on a picture, with the same services. Wan 2.2 can loop the clip back to its first frame.
  • Characters' expressions: the expression generator on a character can make video clips of each expression.

Set the service in Settings › AI Setup › Configuration, on the Video tab.

Local video models are big: plan on a 24 GB card and 64 GB of RAM for LTX-2.5 and MiniMax H3. They stream the part that does not fit the card from RAM, slower but with the same result.

Self-Hosted: Zero-Setup

Self-Hosted: Requires External Software

Cloud-Based

Media processing

Separate services work on a picture or clip once it exists. You reach them from the image viewer's tools (Upscale, Remove Background, Smooth Motion), and from a character's expression generator, which can upscale, cut out and smooth everything it made. Each is set on the Image and Video tabs of the Configuration page. All run locally and only an admin can add them. Each installs on its defaults (Add Modules → Services → name → Install).

ModuleDoesNotes
SpandrelImage and video upscalingSuper-resolution models (ESRGAN, Real-ESRGAN, SPAN, RealPLKSR, SwinIR, HAT, DAT…). Video is upscaled one frame at a time.
InSPyReNetImage and video background removalVideos come out as WebM with an alpha channel. Each frame is cut out on its own, so edges can flicker.
SAM 2.1Video background removalFollows the subject from frame to frame, so its edges hold steady. Experimental.
RIFEFrame interpolationGenerates the frames in between: 2x to 4x the frame rate, same length. Turns a 16 fps Wan clip into 32 fps in seconds.

Voxta Cloud also upscales images and removes their backgrounds, used only when no local module holds the job.

On this page