Skip to content
Bedrova bedrova
FeaturesCompatibilityDevelopersCompareDocs Coming soon

Docs

Video generation

Generate video on your Mac with local text-to-video and image-to-video models.

Bedrova generates video on your Mac. Both text-to-video (describe a shot, get a clip) and image-to-video (take a still you already have and put it in motion) run entirely on-device through MLX on Apple Silicon. Nothing is uploaded, there is no queue, and there are no credits to buy.

This is the newest of Bedrova’s seven capabilities, and the one no other local Mac AI app offers. The tools that do offer video send it to the cloud.

What you need

Video is the heaviest thing Bedrova does. It holds a diffusion model, the frames it is building, and the decoder in memory at the same time.

  • 32 GB of unified memory or more is the realistic starting point.
  • Render time scales with resolution and frame count. Clips are seconds long, not minutes.
  • The pre-load fit-check runs first. If a video model will not fit alongside what you already have loaded, Bedrova refuses rather than pushing your Mac into swap.

See what your Mac can run for the breakdown by RAM tier.

Generating a clip

  1. Open Media, then the Video tab.
  2. Pick a mode: text-to-video, or image-to-video with a source still.
  3. Write your prompt, set frame count and seed if you want reproducibility.
  4. Start the job.

Generation runs as a background job. You can watch progress, keep using the rest of the app, cancel partway, or close the panel and come back. Cancelling releases the memory straight away. Finished clips land in a local gallery alongside the settings that produced them, so a good result is repeatable.

From the API

Video generation is asynchronous on the local API. Submit a request, poll for progress, then collect the finished clip.

curl localhost:1337/api/v1/video/generations \
  -H "content-type: application/json" \
  -d '{ "prompt": "a rover dog trotting through tall grass", "mode": "text_to_video" }'

curl localhost:1337/api/v1/video/generations/<id>

Jobs survive an app restart. See the server and API docs for the full request and response shapes.

Which models are supported

Bedrova ships a reviewed, licence-cleared video pipeline rather than accepting any video repository you point it at. Video model licensing is genuinely messy, and several popular pipelines carry terms that prohibit the kind of redistribution a bundled app requires.

Where a model’s licence does not permit redistribution, Bedrova does not ship, curate or recommend it, and the Model Library says so with a reason instead of failing quietly when you try to load it. That is the same rule the rest of the app follows: a disabled control with an explanation beats a control that looks like it should work.

Memory and safety

Video counts against the same cross-model memory budget as everything else. If you have a chat model pinned and a vision model loaded, the fit-check accounts for both before it lets a video model in. The memory meter in the status bar shows the whole picture, and the Loaded view lets you unload something to make room.