AI in Modular Film Production: Consistency Is the Unsolved Problem
Runway Gen-4, Google Veo 3.1, Lionsgate, Disney, and the Coca-Cola ad case show why AI film production is still a human-heavy workflow in 2026.
Modular filmmaking breaks a production into discrete parts: scripts, voices, visuals, edits. AI tools can now generate each of those parts, but keeping them coherent across a finished film is still the central unsolved problem. Runway and Google both shipped reference-image consistency systems in 2026 to tackle character drift. Studios like Lionsgate, Disney, and Google DeepMind are placing very different bets on where AI fits, and one high-profile partnership, the reported Disney and OpenAI Sora deal, collapsed entirely before it produced anything.
What happened
| Fact | Detail |
|---|---|
| Runway Gen-4 consistency system | Anchors characters, locations, and objects using reference images; Gen-4.5 leads independent video benchmarks on character consistency |
| Google Veo 3.1 release | January 2026; supports up to three reference images to hold a subject steady across clips |
| Lionsgate and Runway equity deal | June 2026; Lionsgate takes equity stake, plans AI short-form series from John Wick and The Hunger Games library |
| Google DeepMind and A24 | June 2026; research partnership and first equity stake in a studio |
| Disney and OpenAI Sora | Reported December 2025; deal collapsed April 2026 after Sora app shutdown; no contract signed, no money exchanged |
| Coca-Cola AI ad campaign (2025) | ~100 people, 5 AI specialists, 70,000+ clips generated; positive sentiment fell from 23.8% pre-launch to 10.2% post-launch |
Modular production divides a film into replaceable pieces. Teams work on individual assets, which can be revised without touching the whole project. That flexibility is exactly why studios like modular pipelines, and exactly why AI complicates them: every generated asset has to match every other generated asset, across different sessions, different prompts, and potentially different model versions.
Runway’s Gen-3 model could not reliably keep a character looking the same from shot to shot. Gen-4 was built specifically to fix that, using a World Consistency system that holds characters and locations steady via reference images rather than retraining. Gen-4.5 reportedly leads current independent benchmarks on this metric. Google released Veo 3.1 in January 2026 with a similar feature: up to three reference images per clip. Two major labs shipping near-identical solutions within months of each other is a clear signal that character drift is the central technical blocker for production-quality AI video, not a niche complaint.
How are studios actually adopting AI in film?
Lionsgate’s approach is the most concrete so far. The studio took an equity stake in Runway in June 2026 and is planning AI-generated short-form series drawn from its own IP, including John Wick and The Hunger Games. According to Lionsgate leadership, the studio sees AI as a creative resource rather than a cost-cutting measure.
Google DeepMind made a parallel move the same month, entering a direct research partnership with A24 and taking an equity stake, marking its first investment in a film studio. Both deals tie the AI provider financially to the studio’s output, which is a different kind of accountability than a pure software licence.
The Disney and OpenAI story went the other direction. In December 2025, a deal was widely reported: Disney would invest over a billion dollars and license more than 200 characters for use with Sora. By March 2026, OpenAI had announced it would shut down the Sora app, citing compute costs and declining engagement. The closure took effect in April 2026. Disney exited the arrangement within days. No formal licence had been signed, and no money had changed hands. Variety, Deadline, and The Hollywood Reporter had all covered the original deal. The follow-through produced nothing.
Why volume alone does not reduce costs
The Coca-Cola 2025 AI holiday ad is probably the most instructive case study available. The campaign required around 100 people, including five AI specialists, to generate and review more than 70,000 clips in order to produce one finished advertisement. Public sentiment analysis recorded positive reactions at 23.8% before the ad launched. After launch, that figure dropped to 10.2%.
That outcome illustrates something important: AI can generate content faster than any previous tool, but without tight review processes, the extra speed converts into more iteration, not less labor. The work does not disappear. It relocates from production to quality control.
For businesses thinking about AI video as a way to cut production budgets, this is the relevant data point. If you are exploring AI integration for content workflows, the Coca-Cola numbers suggest the governance layer needs to be designed before the generation layer is turned on.
Our take
The consistency problem is real and the labs know it. The fact that Runway and Google shipped functionally identical solutions in the same six-month window tells you this is where the competitive pressure is, not on raw visual quality or prompt flexibility. Studios and agencies evaluating AI video tools should treat character consistency across long-form or multi-clip projects as the primary benchmark, not single-shot demo reels.
The Disney and OpenAI situation is a useful reminder that announced AI deals in entertainment are frequently not what they appear. A reported billion-dollar licensing arrangement that produced zero licensed content should be treated as a cautionary example of how much noise surrounds this space.
What the Coca-Cola campaign actually demonstrates is that AI in production is orchestration work. Reference image libraries, prompt versioning, model governance, and approval checkpoints all need to be in place before any speed benefit materialises. That is not a reason to avoid AI video tools. It is a reason to plan the workflow before picking the tool. We cover the broader landscape of AI tools reshaping creative work in our AI news section.
The studios getting traction are the ones that pair AI generation with structured human review, not the ones betting on full automation. That is probably the right frame for most business content production too.
What to do about it
- Test character consistency specifically: generate the same subject across at least five separate clips before committing a tool to a project.
- Build a reference image library and prompt log before starting any multi-clip production. Losing that context between sessions is the fastest way to break visual continuity.
- Assign a review checkpoint after each module, not just at the final assembly stage. The Coca-Cola case shows that catching drift early is cheaper than fixing it in post.
- Track model version updates from your AI video provider. A version upgrade can quietly shift how a locked character or location renders without any obvious warning.
- Set realistic volume expectations: 70,000 generated clips for one finished ad is a real production number, not an outlier.
Frequently asked questions
What is Runway Gen-4's World Consistency system?
Runway Gen-4 introduced a World Consistency system that uses reference images to keep characters, locations, and objects visually stable across different shots, without needing to retrain the model each time. Gen-4.5 leads independent video benchmarks on character consistency.
Did Disney and OpenAI sign an AI deal for Sora?
No formal agreement was ever signed. A deal was widely reported in December 2025, but OpenAI shut down the Sora app in April 2026 citing compute costs and falling engagement. Disney exited within days and no money had changed hands.
How many people did it take to make Coca-Cola's AI holiday ad?
Around 100 people, including five AI specialists, generated and reviewed more than 70,000 clips to produce one finished advertisement for Coca-Cola's 2025 AI holiday campaign.
What is the difference between Runway Gen-4 and Google Veo 3.1 for video consistency?
Both tools use reference images to reduce character drift across clips. Runway Gen-4 uses a World Consistency system anchored to reference images. Google Veo 3.1, released January 2026, allows creators to supply up to three reference images per clip to hold a subject steady across multiple scenes.
