AI is already slashing the cost of commercials, but runaway token fees and piecemeal revisions could leave feature producers paying more than in the days of 35mm film, a Korean AI production pioneer told Taiwan Creative Content Fest (TCCF) in Taipei.

Kim Kwang-jib, founder and CEO of AI content studio Studio Meta-K, said that with the most expensive commercial tools, the token bill alone for a 100-minute film can top KRW500 million ($366,000). AI companies are all losing money, he said, and are keeping prices low as a marketing play.

“If they raise token costs roughly tenfold, it could go above what the films we’re shooting now cost,” Kim said. “KRW500 million could become KRW5 billion ($3.66 million).”

Hollywood, with KRW200 billion ($146.6 million) budgets, could absorb that and still make very cheap AI films, Kim said. Korean, Taiwanese and other Asian features typically shoot live action and VFX for KRW3 billion-KRW5 billion ($2.2 million-$3.66 million), and a tenfold jump would force them to ask whether AI is efficient at all.

He drew a line back to celluloid. Preproduction was meticulous when every take burned through film stock, then digital let crews shoot and cut freely without extra cost. With AI, every re-render is paid for again. If a client sends 10 changes at once, Kim said, Studio Meta-K can make them all and re-render a single time, paying for just one extra pass. If the same 10 changes arrive one a day over 10 days, the studio has to re-render 10 times and pay 10 times over. The final budget, he said, comes down to how few rounds of revisions there are.

“Unless the preproduction stage is made solid, I think AI could end up spending even more money than in the 35mm film era,” he said.

Lee Sang-wook, head of the AI Content Lab at MBC C&I, a subsidiary of Korean broadcaster MBC, said budgeting is harder still because the tools keep moving. His lab’s 139-minute feature “Raphael” took around a year and four months to make, and new tools and updates arrive weekly.

“If you apply technology from a year ago, the scenes look very old,” Lee said.

He added that a single shot on “Raphael” sometimes tied up around 10 artists for an entire day. The film was made with emerging directors discovered by the lab, who raised investment and set up their own studio, and the lab’s work drew backing from Korea’s Ministry of Culture, Sports and Tourism and the Korea Creative Content Agency. “Raphael” will open in Korean cinemas on Nov. 25.

The lab began experimenting with generative AI shorts in 2024, before packaging AI content as a genre strand, “The Prompt: Next Drama,” on Korean streamer Wavve.

“When we made AI short films in 2024, the market reaction was very cold, in Korea as well,” Lee said.

He wanted to face audiences directly, he said, and a cinema release can move to streaming later while the reverse order is hard. He also wanted to show that AI studios, many of which work on an outsourcing model, can take commercial bets through IP development.

On commercials, Kim cited a Pizza Hut spot that aired on TV last year, which took about eight months and roughly KRW2 billion ($1.4 million) to produce. A version Studio Meta-K made in-house took about two weeks and around KRW30 million ($22,000), or about 2% of the cost. In advertising, he said, production costs can fall by more than 90%.

Drama is a harder sell. Kim does not think audiences will yet pay to watch fully AI films or series, so he favors a hybrid model with star actors. A scene requiring more than 500 extras, with costumes, hair and makeup to match, could be generated with AI, leaving only the close-ups of the leads to shoot live. The company is producing war scenes with AI for KBS program “History Special.”

His larger pitch was to reverse the pipeline. Rather than calling in AI after a shoot and edit have already eaten the budget, Studio Meta-K is testing building an entire 100-minute film at film-like quality in preproduction, then reshooting only the parts that feel off, such as emotional bust shots. For now perhaps 50 of those 100 minutes would need to be reshot in live action, Kim said, but he expects that eventually nothing will need shooting at all.

Kim founded Studio Meta-K in 2022 as a K-drama and film company alongside well-known drama directors and writers. Its stated aim is to use AI to cut production costs and own IP rather than take outsourced work, and it has completed more than 100 commercial projects. This year the company was selected for two government R&D projects worth about KRW15 billion ($11 million) combined to build SaaS-based platforms, leading a consortium that includes an MBC subsidiary, KBS and MBN. Studio Meta-K raised around KRW3 billion ($2.2 million) in a Series A last year and around KRW10 billion ($7.3 million) this month, and is preparing for an IPO next year.

The company is also turning Im Kwon-taek’s 1993 classic “Seopyeonje” into an animated feature, re-editing the original footage and converting it into animation with AI for a theatrical release next summer. The project, a collaboration with Busan’s Dongseo University, has the director’s involvement. Studio Meta-K has several animated features in production simultaneously.

The two companies split on open-source tools. Kim said free open-source models have improved sharply, and on one animated project Studio Meta-K builds easy shots with open-source tools and reserves paid commercial tools for difficult ones. Lee said MBC C&I steers away from open-source because it cannot verify the transparency of the training data, noting that no AI model yet meets EU AI Act guidelines. The company has also joined C2PA.

Lee said delivery to streamers is tougher than meeting cinema DCP specs, since platforms analyze files mechanically rather than by eye, and AI output still struggles with requirements such as bitrate and color grading. Kim expects issues like color depth and bitrate to be solved soon.

On jobs, Kim said crews of hundreds on a single scene will grow rarer, and roles such as drone operators and set work could largely disappear. Some companies are already keeping preproduction staff and laying off production crews in favor of co-producing with Studio Meta-K, he said.

“I don’t see AI replacing that workforce,” Kim said. “I personally think workers who have learned AI will replace workers who haven’t.”

Lee compared the shift to the arrival of sound, which cost theater musicians their jobs, and the move from film to digital cameras, which raised the editor’s importance. He predicted new roles including previsualization artists who break down scripts into previs, and continuity supervisors charged with keeping locations and characters consistent. One studio in Korea already has more than 50 AI artists working together, he said, and technical directors will matter more as the technology advances. Spatial consistency, he added, is harder to achieve than character consistency.

Actors, Lee said, have warmed to AI. On a daily drama, previs of a dangerous accident scene turned cast skepticism into approval once performers saw AI would handle shots that could injure them. He does not believe AI can ultimately replace actors’ emotional detail.

“If the ones consuming my content were AI, I don’t know, but the viewers and audiences who consume it are still people,” Lee said.

Archin Yen, director and VFX supervisor at WWWIND Studio, moderated the session. Asked how directors can keep AI from flattening their vision, Kim said AI tools can be learned in a few months, while film grammar such as the 180-degree rule, lenses and editing can take years. Most AI videos he sees default to eye-level, front-on bust shots that film and TV rarely use, he said.

“Every VFX company in the world uses the same tools, but I think it differs depending on who uses them,” he said. “In the end, I think that human 5% decides the quality.”