On a film set the camera move is committed during the take. That has always been the deal, and the entire apparatus of shot lists, blocking, coverage and previs exists because of it — you decide beforehand, because afterwards you cannot.
Generative video reshooting breaks the deal. You can change the move later. But a different move sees things the camera never recorded, and that content has to be hallucinated — and the gap is typically discovered weeks later, in post, long after the location was struck.
Hybrid Cinematography: Previsualizing and Managing Hallucination Risk in Generative Video Reshooting, posted 8 October 2026 by Nhan (Nathan) Tran, Neal Wadhwa, Abe Davis and Stefan Stojanov, puts the discovery back on set.
The mechanism
Two artefacts and one number:
- An editable 3D shot plan — the intended camera move as geometry, not as a storyboard frame
- A proxy of the take — a reconstruction of what was actually captured
- A real-time evaluation of hallucination risk between them
Put a proposed move into the plan and the system tells you where the take lacks support for it. Not “this will look bad,” but specifically which parts of the new view have no recorded basis.
With that visible, the filmmaker has four options the paper names explicitly:
- adjust the plan so the move stays inside what was captured
- explore moves that balance capture and generation — accept invention in some regions, not others
- shoot guided pickups to fill the gaps, while the set still exists
- knowingly accept hallucination
That last one is not a failure case. It is the point.
Why “knowingly” is the operative word
The difference between a tool and a liability here is entirely about when you find out.
Hallucination in a reshoot is not inherently a problem. Plates have been extended, sets have been painted in, and backgrounds have been invented for a century; a director who decides that the wall behind the actor can be synthesised is making an ordinary creative call. What makes generative reshooting hazardous is that the decision gets made by default and invisibly — you request a move, the model complies, and whether it invented the left third of the frame is something you discover by looking at the result and wondering why it feels wrong.
Previsualizing risk converts an invisible default into an explicit choice. The same reshoot, same output, entirely different production posture: one is a surprise in post, the other is a decision in the tent with the DP standing there.
And the guided pickup is the option that only exists on set. Ten minutes of extra coverage aimed at exactly the region the plan is short on costs almost nothing while the crew is standing in it, and is unobtainable once they are not. A risk map that arrives in post can only tell you what you should have shot.
The AR app is the right delivery
The workflow is demonstrated as a mobile augmented reality application for on-set planning, capture and review, plus an offline pipeline for existing video.
AR on a phone or tablet is the correct form factor for this specific job, for an unglamorous reason: the shot plan is spatial and the set is right there. A previs on a laptop requires you to translate between a screen-space model and the room; AR puts the proposed move in the room, so a director can walk the camera path and a DP can see what the move will and will not cover from where they are standing. It also means no cart, no tent, no setup — which determines whether a tool gets used during a pressured hour or ignored.
The offline pipeline matters for a different reason: it applies the same analysis to footage already in the can. You cannot shoot a pickup for a film that wrapped, but you can find out which proposed moves are safe before committing a sequence to a reshoot.
What the study found
A study with experienced filmmakers reveals how previsualizing risk informs camera decisions and exposes tensions between creative intent and generative hallucination.
That second clause is the honest part, and it is the finding worth watching. The tension is predictable and real: the best version of the move is frequently the one with no coverage. The push that reveals the room, the crane that gets behind the actor, the move that needed the wall to not be there — the shots that justify reshooting at all tend to be the ones furthest from what the camera saw. A risk map does not resolve that. It makes you choose, which is more uncomfortable than not knowing and more likely to produce a defensible result.
Filed cs.HC with cs.CV.