A 90-minute film built from 13,000 discarded possibilities
Steelay, a Korean sci-fi action film, is now described as the longest Korean AI feature film to date at 90 minutes, outpacing earlier 60-minute AI movies like I'm Popo and Man in Hanbok. It was made by a production team of just over 10 people in about six months, and it reportedly drew around 5,300 viewers through October 30, making it the most-watched AI feature film released in South Korea so far.
The headline number isn't the runtime, though. It's the 13,000 video clips the team generated and selected from, discarding even more along the way, to assemble a single feature. For a team that size, that ratio of raw output to final footage is the real story, and it says something important about where the effort in AI production actually goes.
For creators building short-form episodic content, this is the detail worth sitting with. The hard part of Steelay wasn't getting AI to generate footage. It was figuring out which of thousands of clips actually belonged in the film.
Generation is cheap, curation is the job
If a 10-person team generated roughly 13,000 clips to land on a 90-minute cut, the math implies an enormous amount of time spent watching, comparing, and rejecting footage for every minute that made it to the screen. The brief doesn't give an exact hours breakdown, but the ratio alone tells creators where their budget and schedule should actually go.
Most creators planning an AI-driven short or episodic series tend to budget time the way they would for traditional production: a chunk for scripting, a chunk for generation, a chunk for editing. Steelay suggests that model understates the real cost center. The generation step can produce output fast and cheap. The selection step, deciding what's usable, what's consistent, what serves the story, is where the actual production time disappears.
The Korea Times review of the film made a related point: it noted visible AI-generated elements but said human selection and finishing work were what delivered a complete commercial film. Director Seol Woo-in was quoted saying that story, characters, action and emotion are what matter most to audiences, not the novelty of the AI process itself. That's a curation philosophy, not a generation philosophy.
- Budget more hours for reviewing and selecting clips than for prompting and generating them
- Treat each generation batch as a casting call, not a finished scene
- Build a simple rating or tagging system early so thousands of clips don't become unmanageable
Tool volatility is a production risk, not a technical footnote
Director Seol told Korea JoongAng Daily that uncertainty around updates to overseas AI tools was the production team's biggest challenge, since changes to the tools or their credit policies could directly affect schedule and budget. The specific example he gave is a useful warning: a new version of the generation tool was released after the team had already finished about 90 percent of the film, forcing them to retest the tool and make additional revisions near the end of production.
That's a mid-production model update undoing finished work, on a project that had already invested roughly six months and a team of 10. For a creator making a single short, a surprise model update is an annoyance. For a creator building a recurring episodic series meant to maintain a consistent look and cast over weeks or months, it's a structural risk that can quietly break continuity between episodes.
The practical takeaway is to treat tool versioning the way you'd treat any other dependency in a production pipeline: something to lock down once you're far enough into a project. If a workflow update could change how your AI characters look, move, or sound, that update needs to happen before principal production starts, not in the middle of it.
- Lock your generation tool version once a project is past early testing, if the platform allows it
- Document the exact model version, settings, and prompts used for your recurring characters
- Build in a buffer in your schedule for re-testing if a forced update happens mid-project
Keeping AI actors consistent across episodes
One detail from the e-Daily interview stands out for anyone building a recurring cast rather than a single film: Director Seol and the production team personally participated in post-recording, or ADR, to breathe life into the AI characters. In other words, even in a film built almost entirely from AI-generated visuals, the voices were handled with direct human involvement.
This matters for episodic creators specifically. Visual generation tools will keep changing versions, and a character's face or movement style can shift subtly between batches or model updates. A consistent human voice performance is one of the few elements a creator can fully control regardless of what the generation tool does next. It's also one of the cheapest ways to maintain the sense of a stable, recurring character across multiple episodes even if the underlying visual model evolves.
Seol has also disclosed plans to develop AI actor stars and produce a sequel next year, which signals that the team is thinking about Steelay's cast as a durable asset, not a one-off generation exercise. That's the same mindset episodic creators need: an AI actor is only useful as a recurring cast member if there's a deliberate system for keeping it recognizable from one episode to the next, and voice work is a practical place to start.
Budget reality: AI tools were still the biggest line item
It's worth being plain about cost, since AI production is sometimes framed as inherently cheap. Ham Seul-gi, head of AI Know Studio, which co-produced and handled AI production on Steelay, said around 500 million to 600 million won, more than half of the film's total budget, was spent on generative AI tools. That's a significant AI tooling spend relative to the rest of the budget, even with a lean 10-person team and no conventional computer graphics.
The brief doesn't specify exactly how that figure breaks down between generation credits, revisions, and the retesting forced by the mid-production tool update, so it's hard to say precisely how much of that cost came from curation-related rework versus raw output. But combined with the 13,000-clip detail, it reinforces the same point: AI tooling costs and selection time, not raw idea generation, are where an episodic production's budget is most likely to get squeezed.
For creators working at a smaller scale than a feature film, the proportional lesson still holds. If more than half your production cost is AI tooling, and most of your generated output ends up discarded, then the planning question isn't how to generate more, it's how to generate smarter and select faster. That's the workflow discipline that separates a usable recurring cast from a pile of one-off clips, and it's the same discipline Episodo creators can apply when building episodic series meant to hold together over time.