Camera and character control across AI video shots stays unsolved

The picture

Three papers on arXiv, spanning December 2023 to December 2025, describe the same unresolved task: getting an AI video model to hold a camera move, a character's identity, or a scene's geometry steady across more than one generated shot. MotionCtrl (6 December 2023) built a controller to separate camera from object motion because prior methods 'either mainly focus on one type of motion or do not clearly distinguish between the two.' ConsistI2V (6 February 2024) calls the problem 'a grand challenge in I2V generation' - preserving 'the integrity of the subject, background, and style from the first frame' across a sequence. Map2Video (19 December 2025), evaluated with 12 filmmakers, reports 'clips fail to match characters and backgrounds' and names 'challenges in shot composition, character motion, and camera control' from its study.

What the documents show

Each paper proposes a technique aimed at the gap, not a claim of closing it. MotionCtrl's language is that its architecture offers more 'fine-grained motion control' than earlier work, not that motion is fully controllable. ConsistI2V proposes first-frame attention and noise initialization, reporting improvement on its own benchmark, I2V-Bench - a measured result, not an external evaluation. Map2Video, the most previs-adjacent, integrates Unity and ComfyUI with a named video model and reports, against its own baseline, 'higher spatial accuracy' and 'less cognitive effort.'

What it is allowed to decide

None of these documents supports more than a class 0-1 use - inspiration or team alignment - for cross-shot planning. None claims, and none of the three papers' language would support a reader claiming, that camera or character control across shots is solved; each names it as the open problem its method targets. None holds Optical authority (matching a real lens across shots) or Provenance authority (each frame's relationship to a specific camera or character asset) at production grade. The papers establish a trajectory: separating motion types (2023), first-frame consistency (2024), filmmaker-evaluated accuracy against a real map (2025) - still framed as narrowing a gap, not closing it.

The disclosure label

A disclosure label for a shot generated with any of these techniques, dated 16 September 2026, would read: this is AI-generated video using a published camera or consistency-control method, decision class 0-1 (inspiration or alignment only), holding no demonstrated Optical or Provenance authority across shots - the cited papers describe cross-shot control as their own open target, not an achieved result. A researcher or tool developer citing their own benchmark is the plausible party asserting this label.

• Does the cited paper claim to solve cross-shot control, or name that as the problem it is narrowing?
• Was reported improvement measured on the paper's own benchmark, or an independent one?
• If a production relies on generated continuity, has anyone checked it against the previs blocking it was conditioned on?

This is an editorial reading of a fast-moving research line: specific numbers will likely be superseded, but the shape of the problem - one clip can look convincing while a sequence cannot yet be trusted to hold a camera move or a face steady - is what a production team should carry forward.