Runway AIF 2026 Seoul Screening: AI Video Is Starting to Truly Become Cinema

Today I attended the Runway AIF 2026: Seoul special screening at the Seoul Film Center. It was a chance to see all ten finalist works from Runway AIF 2026, a leading global AI video festival, in one sitting. The screening ran from 3:00 to 4:30 PM on August 7, 2026 at the Seoul Film Center. The event was organized jointly by AI Contents Lab, MBC C&I, and 100 Going.

All ten works screened were short pieces under ten minutes. That brevity was a limitation, but it also made clear what each film chose to include and what it chose to leave out. According to Runway's official AIF 2026 page, the Film category finalists are as follows. https://aif.runwayml.com/

• A Face Only A Mother Could Love — Robert Gaudette, 7 min 50 sec
• THE WELL — Dorian & Daniel, 5 min 29 sec
• Where Knights Fall — Mathery
• Costa Verde — Léo Cannone & New Forest Film
• TAIRELL ISN'T REAL — Dave Clark, 5 min 4 sec
• Little Mes — Lucas Levitan & Fabián Jiménez, 4 min 19 sec
• Between Before and After — Sifang Chen & Yang Wang, 8 min 21 sec
• POSTMAN — YUUUKI, 7 min
• 살(煞) - Divine Retribution — Minju Choi, 7 min 39 sec
• Simiyu, Lies and Happiness — John Kirk Kelly

Character consistency and simply finishing a film are no longer the question

The first thing I noticed was that maintaining a character's consistency, or completing a work from start to finish, is no longer the audience's first concern. Just a year or two ago, faces, costumes, and spaces shifting from scene to scene in generative video would have been a major obstacle to enjoying a piece. This time, instead of counting technical flaws, I found myself watching how one scene led to the next, what characters were feeling, and why the camera was positioned where it was.

This shift doesn't simply mean the image quality improved. It means the video has begun to feel like cinema. Generative video has moved past the stage of holding the audience's attention through sheer surprise at the output, and has started pulling them along through the timing and rhythm of scenes.

Editing now matters most

What stood out most in this screening was the importance of editing. What shaped the impression of each work wasn't whether AI could generate a single scene, but which scenes were chosen among many, in what order they were arranged, and where the cut landed. The under-ten-minute limit made this even clearer. A continuous stream of beautiful images alone doesn't make a film. It's the gaps between images, the length of the cuts, and the interplay of sound and silence that build a story.

Creators now need to be more than people writing prompts for a generative model — they need to design rhythm in the editing room. There is still a huge threshold called editing standing between having many good shots and having a good film.

The most beautiful works caught the eye first, but the ones with their own worldview stayed longest

Given the ten-minute limit, visually striking works caught my eye first. Since a world has to be established quickly, films with strong color, texture, lighting, and movement grab the audience's attention faster. But looking back after the screening, as a cinephile I felt a certain disappointment in works that skillfully followed the paradigms and mise-en-scène of established film auteurs.

Reproducing familiar cinematic grammar well is different from making a film that is truly one's own. What I found more encouraging in this lineup were the moments that pushed a distinct sensibility and worldview rather than replicating the surface of legacy works. The next stage of AI video will open not through the question of "who did this imitate," but through the question of "is this a world only this director could make."

A summer preserved in the memory of a camcorder: Costa Verde

Among the screened works, one that anyone who has used a camcorder firsthand would recognize was Costa Verde. According to Runway's official description, the film is set in Corsica in the 1990s and follows a boy who gets his first camera and uses it to document a summer with his grandparents. When the screen showed the grandfather and grandmother standing side by side, what the film was trying to do became clear in a single shot. The texture reminiscent of VHS and 16mm home movies, the whirring sound suggestive of a zoom lens turning, and the shifting focal points went beyond a simple retro filter and functioned like an actual memory captured on a camcorder.

That effect wasn't merely decorative nostalgia. It turned the act of holding a camera and watching someone's life into a kind of fantasy. Even though the scenes were AI-generated, they produced the sensation of memory — as if someone really had been holding a camera that summer.

A prince sliding outside the fairy tale: Where Knights Fall

Another work that left a strong impression was Where Knights Fall. Its worldview and mise-en-scène ideas were especially good. Runway's official description explains that Rapunzel waits at the top of a tower while, at its base, the bodies of knights, princesses, and princes who have previously failed the challenge pile up — and finally a man arrives but, instead of climbing the tower, crawls and slips.

It's a parody that inverts the familiar fairy-tale structure of "a long-haired princess waits in a tower for a prince, and the prince comes to save her." What makes this work interesting is that it turns the grand heroism of the original narrative into a farcical, physical chain of failures. It also shows that AI moves closer to its own voice not when it merely reproduces the surface of classic fantasy, but when it twists that grammar.

What comes next: subtle emotional performance and deliberate camera intent

Going forward, results will likely improve the more specifically prompts incorporate an actor's subtle emotional performance and camera effects. It's no longer enough for a face to simply not fall apart. Tiny shifts — a hesitation right before speaking, the speed of an averted gaze, an expression that tries and fails to hide emotion — are what change the meaning of a scene.

The camera works the same way. Whether the same character is viewed from a distance with a long lens, followed with a handheld camera, or isolated in a fixed wide shot all changes the sense of distance the audience feels. Generative prompts now need to move beyond ordering up a "beautiful scene," and closer to directorial notes that specify the relationship between actor, camera, and edit.

A new chapter opening for independent filmmakers

Overall, things have clearly changed from a year or two ago. After watching this screening, I came away thinking that AI video is no longer a genre defined simply by the fact that it was "made with AI." It has already begun to feel like cinema.

That shift could open a new chapter for independent filmmakers. Worlds that used to remain locked in someone's head, because of the cost of large-scale production, equipment, or location scouting, can now be tested as actual footage. Of course, as the barrier to production falls, the bar for direction rises instead. Filmmakers now have to explain not what they can make, but why it should be made with this particular shot and this particular rhythm.

Today's conclusion is simple. AI video has moved from a stage of proving what can be produced to a stage of proving cinematic choices. The works that endure will not be the ones that most convincingly copy a master's mise-en-scène, but the ones that leave audiences with new memories through their own sensibility and worldview.

A pattern only visible after watching all ten together

The ten films used different genres and visual grammars, but they shared a common tendency: hiding AI within the scene rather than putting it front and center. Instead of constantly signaling to the audience "this scene was made with AI," they tried first to convey a character's actions, the mood of a space, and anticipation for the next cut. Technology was becoming part of the production process rather than the headline subject.

This point matters a great deal. In early AI video works, the traces of the generative tool were themselves the spectacle — the way the image warped, unnatural movement, the moment an image suddenly morphed into another form drew the audience's attention. This time, the opposite was true: works lasted longer in the mind when such devices didn't interrupt the story. Audiences no longer checked every frame for signs of generation, and instead followed what the scene was trying to say.

The shorter the runtime, the more a director's approach shows

The under-ten-minute limit wasn't just a screening condition. It was an editorial condition that revealed what each director chose to give up. They had to decide whether to explain a character's past through long dialogue or leave it in a single expression, and whether to explain a worldview across multiple scenes or imply it through the texture of one.

In a short work, a single scene has to do more than one job. It has to introduce a space while showing a character, and explain a character's emotion while setting up the next event. So even a work with many beautiful images can feel fragmented if the relationships between scenes are weak. Conversely, one or two strong images connected to the direction of the story leave a much longer-lasting impression.

This is also why editing felt most important to me in this screening. The ability to generate a good scene with AI is becoming a basic skill shared across many directors. The difference now lies in when to reveal a scene and when to withhold it, how long to hold it, and what sound to pair it with.

Different cinematic approaches across the works

A Face Only A Mother Could Love and THE WELL stood out for how clearly they established character and event within a short runtime. One centers on appearance, gaze, and a mismatch in relationships, while the other builds its world through a fairy-tale-like event and mystery. Both leaned toward making the audience curious about what happens next, rather than simply listing the details of generated imagery.

TAIRELL ISN'T REAL and Little Mes more directly address questions of identity and choice. A character whose very reality is doubted in the online world, and a character who wonders who they might have become had they made different choices, both resonate naturally with "generated images" as a theme. But what matters here is not merely the use of AI as subject matter, but the attempt to turn that subject into a drama of anxiety and regret.

Between Before and After and Simiyu, Lies and Happiness read as works dealing with time, memory, and the distance between people. In particular, rather than explaining the full history of a relationship within a short runtime, they tried to capture a single moment after much has already happened. In works like these, an actor's small expressions, the direction of their gaze, and the timing of a scene's end become all the more important.

POSTMAN made a striking connection between the vanishing culture of letter-writing and emotional recovery, while 살(煞) - Divine Retribution brought a distinctly Korean subject into the world of generative video, through a Joseon-era village, an unfamiliar presence, and the fear shared by a mother, a shaman, and the villagers. What made director Minju Choi's work especially meaningful wasn't simply that it used a Korean setting. It was the way it pushed forward how fear and belief spread within a community, through visual atmosphere.

Beauty is the entry point; worldview is the afterimage

To be honest, right after the screening, the visually beautiful works were the first to come to mind. Color, lighting, the scale of a space, and the smoothness of movement can win over an audience quickly. In works under ten minutes especially, visual polish is essentially the first impression.

But thinking back on it later, beauty alone isn't enough. No matter how refined an image is, if the rules of the world it inhabits aren't visible, the experience fades quickly. Conversely, even if a scene's texture is a bit rough, if it contains actions and emotions possible only within that particular world, the work stays with you.

That's why, in this lineup, I found more potential in works that pushed their own sensibility rather than those that skillfully reproduced an established auteur's mise-en-scène. A work that lets you identify its reference points explains its starting point, but what matters more is what it built anew from there. For AI not to remain a tool that merely redraws past films, a director's personal memory, peculiar tastes, and worldview that resists easy explanation all need to survive on screen.

Prompts now need to direct the camera

Going forward, generative video prompts need to go beyond simply describing a subject and a background. Even with the same character standing in the same room, the emotional tone of a scene changes completely depending on how the camera looks at that character.

A long lens creates a compressed distance between character and audience, while a wide lens shows the structure of a space and a character's isolation at the same time. A handheld camera creates instability and immediacy, while a fixed camera emphasizes a character's inability to escape. A slow push gradually tightens emotion, while a sudden pan forcibly redirects the audience's gaze toward an event.

The same goes for an actor's performance. Rather than writing "a sad expression," it's more cinematic to specify a sequence of actions — closing the lips to hold back tears, looking away, and then eventually looking back at the other person. Emotion arises not from a single expression, but over time.

What independent filmmakers need isn't more images

The idea that AI will open a new chapter for independent filmmakers became clearer after this screening. But that chapter doesn't simply mean an opportunity to produce more scenes at lower cost. It's closer to meaning that creators can now quickly test scenes that used to exist only in their heads, discard directions that fail, and try different editing structures.

In the past, filming a single hypothesis required renting a location, gathering actors and crew, and preparing equipment. Now that hypothesis can be checked as footage first. But that also means more judgment calls are required. Directors must decide whether to accept a generated scene as is, redesign it as if it were actually filmed, or use a scene's flaws as a deliberate style.

Ultimately, the competitive edge independent filmmakers need isn't knowing the generative tools best — it's the instinct to choose what to make. Generative video lowers the barrier to production, but it doesn't lower the bar for what counts as a good work. In fact, the less audiences worry about technical flaws, the more strictly directors will be questioned about story and direction.

What remained after seeing it firsthand in Seoul

This screening wasn't a place to ask whether AI video can become cinema. It was a place to ask what works that already feel like cinema still need. The answer was editing, emotional performance, camera intent, and a worldview of one's own.

In AI video from a year or two ago, the mere fact of having been made was itself the event. Now the event comes after that. Audiences start asking which scene a director chose among countless generated results, which scenes were discarded, why this character stays silent at this moment, and why the camera stops here.

Works that can withstand that question are the ones that will survive going forward. Advances in technology are already giving many people a similar starting line. The way to move far beyond that starting line is not to pile up flashier images, but to hold onto one's own rhythm, sensibility, and worldview to the very end.

This piece is a firsthand account written after attending the Runway AIF 2026: Seoul special screening held at the Seoul Film Center on August 7, 2026. Titles, directors, runtimes, and official synopses were cross-checked against the official Runway AIF 2026 page.