AI video's consistency problem, and what Google's new research changes

On 24 September 2026 Google Research published a multi-agent framework for long-form AI video and named the two failures it exists to prevent: identity drift, where a character's face, clothing or a room changes quietly between shots, and cascading failure, where one bad asset corrupts everything generated after it. Both are created before any animation happens, which is why an approved still is cheaper than a re-render. The same day, Adobe moved Photoshop, Lightroom, Express and Firefly into Gemini, and Acrobat into Claude. Read as a diagnosis rather than a release, the paper says the consistency problem is a decision problem — and it describes the film-makers who write shot lists and lock their look as the ones who already have it solved.

The paper introduces a unified multi-agent framework that generates temporally consistent, long-form video narratives on its own. The interesting part is not the output — it is the diagnosis. Google names two failure modes in the pipelines people use today: semantic drift, described as subtle shifts in character attire or scenery across shots, and cascading failures, where an upstream asset artifact corrupts the video synthesis downstream.

It also names the cause. Current pipelines chain modules together with independent, handcrafted prompting, so early errors propagate and break consistency over a long horizon, which is why the process ends up needing exhaustive manual intervention. That is a description of every AI film-maker's week.

Identity drift is the face that changes between two shots of the same person. Cascading failure is the corrupted frame you do not notice until it has been carried into six clips. Anyone who has made a character survive more than a few shots has met both, and neither is fixed by a better model.

What fixes them is deciding things once and then holding them: who the character is, written down as fields rather than adjectives; how the film looks, approved as a single reference sheet; and which still is allowed to become motion. Those are decisions, not settings.

The framework's central insight is structural: because errors propagate forward, the cheapest place to fix one is upstream. That is the same argument for approving a keyframe before spending anything on animation — a still costs a fraction of a clip, and it is the last point where a wrong face can be corrected for free.

Read that way, the research is not a reason to wait for better tools. It is a description of why the people who plan their shots get consistent films out of the tools that already exist.

Also on 24 September, Adobe brought Photoshop, Lightroom, Express and Firefly into Google Gemini, putting them in front of more than a billion monthly users, and added Acrobat to Claude. Adobe already sits inside ChatGPT, Microsoft Copilot and Slack. The company's own framing is that people will increasingly start creative work inside an assistant rather than inside Adobe's applications.

That is a bigger change than it sounds. When the tool is a sentence in a chat window, the limiting skill stops being where the buttons are and becomes what you asked for, in what order, and what you refused to accept. The same week, GPT Image 2.5 — the model behind the stills in my own pipeline — arrived inside Adobe Firefly.

Access is no longer the dividing line here. The assistants are on every phone, and the Singapore workforce is among the world's more active users of AI at work. At company level the picture is different: the Ministry of Manpower's first AI adoption report, published on 30 April 2026, found 71.5 per cent of firms had not adopted AI at all, rising from 23.9 per cent adoption among firms with fewer than 25 employees to 76.4 per cent among those with more than 500.

Read together, those two facts are encouraging for small studios and small brands. The tools have been democratised; the method has not. A two-person shop that writes its shot list, locks its look and approves its stills is competing on the part that is still scarce.

Write the shot list before generating anything: what each shot is for, how long it lasts, and what must not change in it. Lock the look once as a single approved reference and attach it to every shot rather than describing it again. Then approve a still before you animate, because that is the point where a drift costs you nothing to correct.

None of that depends on the next model release, which is exactly why it is worth learning now.

The interesting part of the Google paper is the diagnosis, not the framework. A multi-agent system that holds a film together on its own is not something a two-person studio will be running this year, and we would not plan a project around it arriving soon. What we would take from it is the confirmation: errors propagate forward, so the cheapest place to fix one is upstream — which is exactly what approving a keyframe before you animate already does.

The Adobe news is the one we would watch. When the tools live in a chat window, the craft stops being where the buttons are and becomes the brief: what you asked for, in what order, and what you refused to accept. We are less convinced than the industry that a better model settles identity drift. It has always looked like a decisions problem to us, and this is the first time a large lab has said so on the record.