Netflix Details AI Use on 300 Titles at Busan as Kang Yoon-sung Cites a 40% Budget Cut

TL;DR

On October 10 at the Busan International Film Festival, Netflix's senior director of APAC production, Lee Sung-gyun, told a panel that AI has been used "in various ways" across roughly 300 Netflix titles: previs, production design, set-dressing concepts, editorial and VFX. That matches the number in Netflix's Q2 2026 shareholder letter, but the panel is the first time the company has walked through where in the pipeline the tools actually land. Sharing the stage, The Outlaws director Kang Yoon-sung said his fundraising math for the AI-built naval thriller The Gulf of Aden came out about 40% below a conventional budget, and Wētā FX CTO Kimball Thurston drew a hard line between "assistive AI" that fixes pixels and generative tools that invent them. The whole panel agreed on one caveat: the footage looks great and the emotion got worse.

What Netflix actually said

The session, "Film Production and AI: New Tools, New Possibilities," was co-hosted by Netflix and the MPA as part of Creative Asia, the Netflix and BIFF program for emerging Asian talent. Variety and The Hollywood Reporter both carried the quotes, with the usual Korean romanization drift (THR calls him Sung Q Lee and the director Kang Yun-sung).

"Across around 300 Netflix titles, AI has actually been put to use in various ways. It's being used in previs, in production design, in working out set-dressing concepts ahead of time, in editorial, and of course in VFX."

The 300 figure is not new. The Q2 letter, published July 16, said GenAI workflows had been used "in roughly 300 of our titles" in 2026, "with the largest concentration of work in post-production," and named Glory (India), Brasil 70: A Saga do Tri (Brazil) and The American Experiment (US) as productions that used the tools for "enhanced crowds, historical battle sequences, and worldbuilding establishing shots." What Busan added is the stage-by-stage breakdown and a reel from Eyeline, Netflix's VFX studio, on what the company calls hybrid production: live-action photography, conventional VFX and generative tools in one pipeline.

The one concrete example was the Netflix India title Glory, where Lee screened a clip of AI and VFX together producing large crowd scenes. Under the show's budget and schedule, he said, those scenes would normally have been cut. That is the same framing as the shareholder letter: not "we replaced artists," but "shots that would have been dropped got made." You can decide how much of that is accounting and how much is art.

Asked how Netflix keeps up with new models, Lee said the company works under strict internal guidelines and tests tools before adopting them. "If you're asking about speed, I'd like to say that we're not wildly obsessed with it." For a company that just told shareholders GenAI utilization "is scaling quickly," that is a fairly relaxed definition of not obsessed.

Kang Yoon-sung's 40%

The more useful numbers came from Kang, whose The Outlaws franchise is one of Korea's biggest box-office properties. Last October he co-directed Run to the West with AI filmmaker Kwon Han-seul, billed as Korea's first feature to use AI across its entire production. By his own account the tools were crude: "we mainly used it for compositing and making creatures. And back then, even the creatures didn't blend well with the footage we'd shot."

A year later he says commercial use is "no problem at all." His test clip for The Gulf of Aden, a film built around military operations at sea, was made entirely with Kling, the Kuaishou video model, earlier this year, rendered at reduced quality for the test. Current tools, he said, can output 4K at 10-bit color or higher, which clears the bar for cinema delivery.

"When we calculated how much we'd save on 'The Gulf of Aden' for fundraising, it came to an actual reduction of around 40%. That's not just from replacing CG and VFX. The shoot itself got simpler, and it cut the overall production period to some extent."

Read that carefully. It is a fundraising projection for a film that has not shot yet, not a post-mortem on a delivered budget. Earlier in the session he gave the range as 30% to 40%. The savings he describes are structural rather than line-item: fewer CG shots, yes, but also a simpler shoot and a shorter schedule, because sequences that would have needed a sea, a ship and a second unit become generated plates.

The precedent he can point to is Run to the West. Per the Korea Times, it ran 61 minutes, about half a normal feature, was completed in five months, less than half the usual time for an effects-heavy project, and sold tickets at 8,000 won, half the going rate. One car explosion that would take four or five days in CG took about a minute to generate. The deal with its backers was explicit: a second part only if the first drew at least 200,000 admissions. That is the honest shape of AI film economics right now: shorter, cheaper, discounted at the box office, and funded one gate at a time.

Wētā's line: fix pixels, never invent them

Thurston was the most careful speaker on the panel, and the most technical. He described the deep learning denoiser behind his second Academy Scientific and Technical Award (a 2025 Technical Achievement certificate shared with Javor Kalojanov) as "assistive AI."

"It is basically not allowed to change the creative intent in any way. It touches pixels, but it only fixes them."

That distinction is the clearest definition of the two camps you will hear from a working VFX house. A denoiser takes a render that was stopped early (noisy, because the path tracer did not get enough samples) and predicts what the converged image would have looked like, using temporal filtering and optical flow so fine detail survives across frames. Think of it as a spell-checker for a frame: it will fix the typo you made, but it will not rewrite your sentence. A generative tool, by contrast, is a ghostwriter, and the question of whose sentence it is becomes a lot more interesting. Wētā, Thurston said, has used machine learning for at least a decade, or 20 to 25 years if you count early forms of Massive, the crowd system it built for The Lord of the Rings.

He also flagged a failure mode that anyone training on scraped video should write down. "If all your AI models have been trained on video images, you're not getting the right red for a Coca-Cola can. You get Netflix red, but not Coca-Cola." Models that learn color from compressed, display-referred streaming footage inherit that footage's gamut and tone mapping, which is not the scene-referred color a brand or a colorist signs off on. Learning what red looks like from Netflix encodes is like learning what a tomato looks like from photocopies.

On provenance, Thurston said Wētā is building an open-source tool through the Academy Software Foundation to track AI use and lineage inside a production, designed to link with C2PA. If it ships, it is the first piece of pipeline-level AI disclosure tooling from a top-tier VFX vendor, and it would give studios something to point at the next time a guild asks what exactly was generated.

The caveat everyone agreed on

The quotable line was Lee's: "a human touch is absolutely essential." The substantive one was Kang's, and it was not flattering to the tools. Because AI handles so much of the creative work, he said, it often cuts out the discussion, effort and experience of the many people filmmaking has always relied on, leaving images made from one person's ideas alone.

"So the quality of the footage AI makes is very high, but when it comes to delivering emotion and moving audiences, I think we're in a period where it has actually gotten worse."

Kang expects the market to split. Short-form content will optimize for productivity, and he pointed to China, where some short-form projects are already made entirely with AI from planning through distribution. Film and drama will optimize for credibility instead. On actors, both he and Thurston said replacement is unlikely and stunt and stand-in work is where AI doubles land first: "even if AI acting becomes very sophisticated and imitates very precisely, emotion doesn't come through because of its mechanical coldness."

Two housekeeping notes from the sustainability question. Thurston said Wētā's Wellington data center, on closed-loop cooling, now draws "zero liters of water per year." Lee said Netflix looks at sustainability company-wide against its 2030 emissions goal and does not break out AI's share separately, which is a polite way of saying the number is not public.

Why builders should care

• The customer is post-production. Netflix's own letter says the concentration is in post, and the panel's only concrete example (crowds in Glory) is a VFX shot. If you are building creative tooling for studios, the wedge is shot extension, crowd fill and denoising, not "generate my movie."
• Color is a real moat. Thurston's Coca-Cola red complaint is an eval you can run today. A model that cannot hit a brand color in a scene-referred pipeline is not cinema-grade, however pretty the demo.
• Provenance is becoming infrastructure. An ASWF tool linked to C2PA means AI usage metadata will ride alongside shots the way color metadata does. Design for it now.
• The 40% is a pitch deck number. Treat it the way Kang's financiers presumably did: plausible, structural, and unproven until a film delivers on it.

Key Takeaways

• Netflix's APAC production head said AI has been used across roughly 300 titles, in previs, production design, set dressing, editorial and VFX, matching the Q2 2026 shareholder letter.
• The letter puts the largest concentration in post-production and names Glory, Brasil 70 and The American Experiment as productions that used GenAI for crowds, battles and establishing shots.
• Director Kang Yoon-sung says his fundraising model for The Gulf of Aden, test-rendered entirely in Kling, came out about 40% cheaper, with 30-40% quoted as the general range.
• His 2025 precedent Run to the West ran 61 minutes, finished in five months and sold half-price tickets, with a sequel gated on 200,000 admissions.
• Wētā FX draws the line at "assistive AI" that fixes pixels without changing intent, and is building an open-source AI-provenance tool through the Academy Software Foundation to link with C2PA.
• Everyone on stage agreed footage quality is up and emotional impact is down. Kang's words: "it has actually gotten worse."

Sources: Variety, The Hollywood Reporter, Netflix Q2 2026 shareholder letter (PDF), The Korea Times, Wētā FX awards