A brand film showed what a generate-and-reject loop still needs

The process wall

An origin-story brand film set in an earlier era normally needs sets, casting and a VFX budget most retailers cannot justify for a single festival spot. In June 2024, Toys'R'Us Studios and the creative agency Native Foreign, working with early access to OpenAI's then-unreleased Sora, produced a short film for the Cannes Lions festival by generating video directly from text prompts instead of shooting it, replacing that production wall with an iteration problem: getting a generative model to return usable footage at all.

What the documents show

According to reporting on the production, Toys'R'Us global CMO Kim Miller Olko described the approach as 'the 2.0 of what I call "same magic, new method" era,' and said some generated iterations were 'exactly logically what we said' in the brief but 'didn't experientially fit into the human experience like it would in real life,' meaning they were rejected. Native Foreign's chief creative officer Nik Kleverov compared Sora's role to 'a new camera and a post [production] in one,' and said prompts written in filmmaker's terminology produced better results. A separate account adds that the process condensed 'hundreds of iterative shots down to a couple of dozen' over a few weeks.

Authorship and limitation

The mechanism both accounts describe is generate-then-reject, not one-shot prompting: a human standard of what 'fits the human experience' sat over the model's output at every stage, deciding what survived from a much larger set of generations. What neither source documents is disclosure at the point of viewing. The explanations from Olko and Kleverov are process statements given to press after the film was already public, not a label carried on the film itself, and Digiday's reporting notes the finished piece was called 'creepy' by some viewers, a reaction an on-screen disclosure might have pre-empted or at least reframed.

What survives

What survives is a workable description of the craft: iterate broadly, judge against a human standard, keep only what passes. Editorially, the gap this case leaves open is that explaining a process after release answers a different question than disclosing it at the point of viewing, and the two should not be treated as equivalent.

• Separate a brand's after-the-fact process explanation from an on-screen AI disclosure; this case had the former without the latter.
• Ask how many generated iterations were rejected before the version shown, since that ratio is the real measure of the creative loop.
• Note that prompt phrasing drawn from filmmaking vocabulary is reported as improving results, worth treating as a craft skill rather than an incidental detail.

The film is documented mainly through what its makers told journalists afterward, which is useful evidence of process but is not the same record a viewer would have wanted at the moment of watching.