Reconstruct videos as editable projects

Reverse Any Video guides reconstruction of an authorized source video as an editable HyperFrames project. The workflow measures the original's dimensions, frame rate, duration, cuts, motion, and layer placement, then rebuilds the composition rather than playing the original whole frame. Its goal is a render that closely matches the original while keeping reconstructed elements editable.

The original determines the root composition size and frame rate, including fractional rates such as 30000/1001. Duration comes from video frame count divided by fps rather than audio length. Cut measurement includes jump cuts and punch-ins inside a take. Shot start times are calculated as frame divided by fps and rounded down so a decimal approximation does not move a cut one frame late.

Use it when the user explicitly wants an authorized video reconstructed in HyperFrames. Rebuild the video as supplied; remixing requires a separate request. This advanced optional workflow must not automatically trigger paid reconstruction.

Reverse requires a signed-in paid HyperFrames Pro plan. A resumed project must repeat the authorization and startup checks, including any relevant credit limit. A provider's policy refusal cannot be worked around or retried as a reconstruction tactic. Transient failures can be retried only when billing and task authorization remain clear. Rights to footage, likeness consent, the payer, and authorization for paid generation are required before recreating a person with HeyGen.

Reconstruction begins with an authorized original video, a compatible HyperFrames toolchain, and applicable cost limits. Measurement starts with:

npx hyperframes measure probe <video>

Consecutive frames around cuts, entrances, exits, and transitions reveal changes that widely spaced contact sheets miss, including short blurs, dissolves, flashes, and motion direction. Large words are inspected throughout their onscreen lifetime. Text sizing, caption switch frames, backdrops, and visual treatments follow the matching guidance.

Components such as text, captions, stickers, cards, UI, charts, and maps are rebuilt in HTML. Photos, clips, album art, logos, app icons, and collage subjects are extracted into their own boxes. Logos use the largest suitable original crop or the real mark rather than a hand-drawn substitute. The original video is copied into the project as a single audio source across the film, avoiding extracted or re-encoded audio; reconstructed clips remain muted.

During construction, selected moments can be compared without rendering. Once every shot has content, run the HyperFrames checks and correct errors, then render and compare the result against the original. Cloud rendering follows the supported frame shapes and rates. Fractional rates, other frame shapes, or failed cloud renders use the local rendering path, with a foreground timeout or a detached long-render process whose log is observed.

The diff identifies missing or extra cuts, duration discrepancies, sound offsets, missing-content stretches, and poor visual matches. Fix visible issues such as wrong timing, dimensions, weight, or color, then render and compare once more. Small score differences do not justify an endless loop. The final REVERSE.md records rebuilt, extracted, and re-shot material, approximations such as substitute fonts, final diff numbers, and relevant remaining authorized credit limits.

The workflow uses HyperFrames with tools such as ffmpeg and Python, plus HeyGen only for appropriately authorized recreation. Remote providers and hosted services retain separate availability, licensing, and cost constraints. Use the selected CLI version without automatically upgrading it.

Default helper invocations disable telemetry, skill updates, and update checks. External reports require a user request, and credentials stay in existing secret storage. On resume, commands recorded in notes are reconstructed from their original recipes and checked with the planning option rather than executed blindly. Actual visible output and accessibility must be checked before delivery; recipe examples and fidelity scores do not establish success in every environment.