Brand Systems
Client
Bally's INteractive
Timeframe
Ongoing
Year
2026
The situation
Football made continuity obvious. In one sequence, the ball changed appearance between shots and a celebration appeared before the event that caused it. Tennis exposed screen direction. A player served, appeared to lose the point and then celebrated. Another cut crossed the 180-degree line and reversed the direction of play.
Horse racing exposed visual consistency. A rider could feel slightly different between shots, while small branded details moved around the frame. These looked like exclusively gen-AI problems at first, but most of them were familiar production problems. Editing, continuity, casting and screen direction already had established solutions.
That changed our reviews. A generated shot had to work inside the sequence, not simply look convincing on its own. We started paying more attention to cause and effect, screen direction, recurring objects and whether a character really felt like the same person from one shot to the next. Sometimes that meant discarding an expensive generation that looked good by itself because it did not belong in the edit.
Building the missing production structure
We started carrying more decisions outside the generation itself. Character sheets became a useful way to establish people before producing the shots. Animation has used them for decades so that different artists can reproduce the same character consistently, and the same principle works well with generative models.
We applied the same thinking to locations, wardrobe and recurring objects. The reference bank became a way of keeping those decisions stable from shot to shot. It closely resembles the structure of a conventional shoot: you cast the person, choose the clothes and establish the location before moving through the shot list. Those decisions remain available throughout production.
Movement needed a different approach. We could spend a great deal of time explaining a ball bounce, body action or camera move and still receive something that looked physically wrong, so we began using motion references instead. For camera movement, I also used simple 3D previs where useful. A rough 3D animation can communicate framing, timing and trajectory more clearly than a long description.
That became a practical rule in the workflow: when motion is important and language is ambiguous, show the movement.
Control changed how I evaluated models
This work also changed what I value when testing generative video tools. Generation speed and visual fidelity are easy to compare, but production needs another dimension: how much of the result can I intentionally control?
Reference capacity became particularly important. Models able to accept several images, character sheets and video references gave us more ways to preserve decisions across a sequence. A faster model can still be useful for exploration, but for production I care much more about whether a creative decision can be carried from one shot to another with reasonable predictability.
That is also why our workflow gradually became more reference-driven. Information about the cast, location and movement lives in assets we can inspect and reuse instead of depending entirely on what the model remembers from a written description.
What changed
The sports work gradually became less dependent on prompting and more dependent on production discipline. Characters and locations could be established before generation, motion could come from references, camera behaviour could be planned in 3D, and editing reviews included continuity and screen direction alongside image quality.
Each project changed the next one. Football made continuity more visible, tennis sharpened the editorial review, horse racing pushed the reference system further, and other sports continued to stress physics, crowds and movement in different ways. The models still make mistakes, but we became better at understanding where those failures came from and moving more of the important decisions into parts of the workflow we could control.
Reference-driven generation eventually created another bottleneck: preparing all that information for the model became a substantial task of its own. That became the next problem I worked on.
Why this case matters
Most of the useful ideas in this system existed before generative video. Character sheets came from animation, previs came from 3D production, continuity and screen direction came from filmmaking and editing, and compositing gave us control when generation alone did not.
The important part was recognising which production discipline answered the failure we were seeing. Generative video changed the medium. It did not change the fundamentals of making a sequence work.
Discover other projects
Dive into our diverse collection of innovative projects, where creativity meets cutting-edge technology to solve real-world challenges