All articles

Craft

AI Video Repair vs Regeneration

In most generation workflows the smallest thing you can change is the thing you generated. Generate an episode, and the smallest fix is an episode. That single property is why correction costs in AI production compound the way they do.

Close-up of fictional Anaya showing facial detail, wet hair and warm interior lighting
AI-generated concept artwork for Tosheo. Fictional people and scenes; not a customer production.

The blast radius

The costs of a regeneration are three, and only the first is visible on an invoice.

The render itself. The re-review of every shot the run touched, including the ones that were fine. And the variance: a new run is not the previous run, so shots you had accepted now differ, and the difference is either accepted again or becomes a second problem.

The fix is structural rather than clever. Make the unit of generation, the unit of review and the unit of repair the same small thing, and the blast radius collapses.

RegionSmallest visual repairPart of a frame
ShotStandard repair unitReplaced alone
SegmentAudio repairA line, not a track
VariantDerived repairOne version, not the master

Repair depends on reproducibility

A narrow repair is only possible if the conditions of the original can be reproduced: same provider, same parameters, same references, same canon version. That is the practical reason a treatment is locked once approved, and the reason a router that always chases the best available model makes its own repairs harder.

It is also why the record kept around a repair matters as much as the repair. The original, the change, the reason, the approver and which deliverable carries which version all stay in the production - for creative reasons, for the delivery manifest, and because attempts per accepted repair is a cost metric you cannot manage if repairs overwrite their own history.

Fictional exampleEleven frames

Shot 04 is right in every way except that a hand is malformed for eleven frames as it reaches toward a prop.

Wrong unit
A region, across part of one shot
Repaired
That region, same references and treatment
Untouched
Every neighbouring accepted shot
Untouched
The voice track for the scene
Recorded
Original, repair, reason, approver, deliverable

The episode does not re-render, neighbours do not need re-review, and the audio is not re-timed. The cost of the fix ends up close to the cost of the thing that was wrong, which is the only sane relationship between the two.

What cannot be repaired narrowly

There is a second failure worth naming: the repair that fixes the hand and shifts the light. Repair success is measured as success without collateral change, because a fix that quietly alters something else has not succeeded, it has moved the problem.

Frequently asked questions

Can any generation tool do shot-level repair?

Several can regenerate or edit a piece of media. What is usually missing is the surrounding structure: a shot that was generated, reviewed and accepted as a unit, with the references and treatment recorded so the conditions can be reproduced, and lineage that survives the change.

What is collateral change?

Anything that moved outside the region you asked about - lighting, skin tone, a background object, the grade. It is the most common way a repair makes a shot worse while appearing to fix it.

Is the original kept after a repair?

Yes. Approved work is not silently replaced, the original stays in the production record, and the delivery manifest states which version was delivered.

Do repairs come out of my budget?

Repairs needed because a provider failed or returned an unusable result are our cost. Repairs from a change of mind are a scope change and are estimated like one. The distinction is recorded rather than argued about at invoice time.

Have you measured repair performance?

Not publicly yet. Attempts per accepted repair and repair success without collateral change are tracked, and they will be published from our own demonstration production rather than asserted now.

Tosheo

Notes on the craft, decisions and systems behind AI-native production. Published by Tosheo.

Updated
All articles

From reading to making

See how a production comes together.

Follow the AI Director from your story brief through planning, review, production and delivery.

How Tosheo works