Frame — AI Video Generation & Editing

Three ways to make a video

Generate starts with a written description and an optional reference image. The app can develop the prompt, generate a candidate and retain earlier versions for comparison. Highlight and Demo start with uploaded footage: Highlight selects moments for a creative edit; Demo aims to preserve a followable instructional sequence.

For existing footage, the workflow moves through analysis, direction and an edit plan before export. A low-resolution preview is optional. The plan can specify source ranges, ordering, speed, titles and sound; FFmpeg applies those instructions.

Revisions you can inspect

Saved LTX generations retain their exact prepared prompts. Version history lets the user revisit candidates, keep a preferred version and reject later versions without deleting the files. Optional creative notes can be accepted or rejected before a revised prompt is prepared.

Changing the scene regenerates the footage. It does not preserve every face, prop or pixel from the previous version. Changing cuts, timing or titles instead uses the existing-footage editor. That separation helps the user choose the right operation for the requested change.

Local workspace, configured compute

The Windows application uses a Python/FastAPI backend and a browser interface, with local project files and exports. The configured setup uses DGX Spark for video generation and selected AI services. A local workspace therefore does not mean that every model runs on the laptop.

The inspected LTX generation profile supports short clips up to five seconds at 768 × 512, and a single longer clip up to 45 seconds at 512 × 320 on Spark. The examples above are the supplied MP4 exports; their resolution and duration were checked directly.

An experimental Wan2.2 option provides temporary generation separately from saved LTX versions. Its results must be downloaded to keep them; they do not enter project history automatically.

Review and current limits

AI observations and creative suggestions still require judgement. Sampled frames can miss brief events, speech timing can be wrong, and generation may omit a requested action. A completed render does not establish narrative quality, exact prompt adherence or reliable lip sync.

The footage-editing workflow tracks approvals and invalidates downstream approvals when earlier inputs change. This helps prevent an old export being mistaken for a newly approved edit. Frame remains a single-user local application in development, with no built-in public hosting or publishing service.