Egor Kuzmin’s AI video reached stadium screens

Egor Kuzmin’s stadium project makes the case for AI video more cleanly than any polished demo: generative footage can hold a three-minute narrative on giant public screens, provided it enters the process as raw editorial material and gets cut, scored, and checked like a real film. This was no one-prompt spectacle. It was a medieval battle sequence for a live-show interlude, placing an artist as a knight among a castle, soldiers, and horses.

Kuzmin is an AI-video creator and founder of XR School and XR Production. Earlier, he worked in architecture and taught at an architecture school; the Institute of Music Initiatives also credits him with music-video participation for Eldzhey, Onative, Wipo, Artem Shilovets, and Erika Lundmoen. That background earns its keep on a stadium screen. Composition, scale, continuity, pacing, and the emotional read of a silhouette all have to survive distance and motion.

The useful lesson for creative teams is not that every campaign needs a fortress and cavalry. AI generation can now support visual concepts that would usually require a major location, cast, wardrobe department, and VFX pipeline—if a human team owns the brief and the finishing calls.

What did Egor Kuzmin create for the Luzhniki shows?

For two Luzhniki shows in July 2026, Kuzmin made roughly three minutes of a three-minute-thirty between-songs film. His section stages a medieval battle: the musical artist is a knight, backed by a castle and surrounded by a dense field of soldiers and horses. Other creators made the program’s remaining segments.

That split is the right production model for large-format AI video. A show is a run of moments with different visual jobs, not one uniform deliverable. One creator can take a defined narrative passage while editors, audio specialists, show creatives, and other artists shape the full program around it.

The brief had a usable visual grammar. Knight, castle, cavalry, and massed troops give the model repeatable anchors for period, scale, character status, movement direction, and atmosphere. Set comparable anchors before anyone generates: hero subject, environment, action, palette, lens language, duration, and the clip’s place in the final edit. “Make it cinematic” is useless as a stadium brief.

How did a short AI clip become a stadium commission?

A free fifteen-second Seedance 2.0 music snippet set the chain in motion. Three minutes after Kuzmin posted it, a music-video director contacted him; Kuzmin then contributed to the related Malo 2.0 video before the show’s creative team asked him for the stadium-film segment.

Study that route because a finished public proof of taste beats a broad claim that you can do the work. The first clip was short. Still, it showed a visual point of view quickly enough for a director to place it in a music-video setting. For a brand or agency, a thirty-second proof sequence should show the exact hard thing the buyer needs: a reflective product rotating under controlled light, believable apparel drape in motion, or an environment that stays coherent across several cuts.

Fast discovery does not remove production planning. It changes the first gate. A compact proof earns the conversation; the shot list, continuity rules, review checkpoints, and final technical delivery plan win the commission.

Why did the project require fifty-five generated videos?

Those fifty-five clips expose the real unit of AI-video work: selection. Kuzmin generated his portion in Seedance 2.0 at 4K, then an editor assembled it into the finished segment and a sound engineer mixed the final track. Model output is candidate footage; the film comes from choosing shots that work together and cutting them against music.

Action scenes make this obvious. A cavalry charge may convince in one burst, then lose its direction in the next; a castle may change architecture from shot to shot. The knight’s costume, weapon, face, and scale against the crowd all need a deliberate continuity pass. Generate shot families instead—establishing view, hero close-up, lateral action, overhead massing, exit beat—rather than asking one prompt to produce a whole narrative.

Generation speed gives you an iteration budget, not a publication deadline. Lamina’s current median asset-generation time is 147 seconds, and its ninety-percentile time is 493 seconds. Those numbers let a creative lead reserve time for parallel exploration, while excluding prompt development, brand review, revisions, edit assembly, sound, and delivery checks. The approved asset always arrives later than the first generated one.

What does a stadium screen change about AI-video direction?

A stadium screen raises the creative bar: every shot has to read instantly from far away. Kuzmin’s battle concept holds at that scale because its major shapes are plain—a mounted knight, a fortress, a moving army, high-contrast action. Fine prop detail can enrich a frame. It cannot carry the story alone.

Direct with hierarchy. Put the hero action in the strongest part of the composition, limit each shot to one dominant event, make transitions legible on the beat, and repeat visual markers so viewers can find the protagonist after a cut. Product video follows the same hard rules: show the SKU early, hold its color and material steady, give it enough screen area, and avoid transitions that bury its functional detail.

Four-kilobyte-resolution delivery is only part of the discipline. A 4K-generated source gives the finishing team room to cut for a large display, though a technically large file cannot rescue muddled blocking. Review exports at the intended aspect ratio, in a setup that approximates the real viewing distance, before you lock the sequence.

How should a team produce AI video for a live event?

Begin a live-event AI-video workflow with a locked narrative beat sheet and finish with human technical review. Kuzmin’s project had a bounded role—a between-songs segment—and a named handoff to an editor and sound engineer. Generating isolated scenes and hoping to find the story in the final render window is a bad way to ship.

Break the piece into shots, each with one visual objective. Define the hero subject, setting, action, framing, transition, and duration for every shot. Then make multiple candidates for the beats most likely to fail: faces, hands, fast movement, brand marks, material close-ups, crowd geometry, and any shot that must connect directly to the next.

Build a select reel before you commit to the edit. Review continuity, brand appropriateness, unwanted visual artifacts, readable action, and whether changes in the music require a new cut point. Keep approved reference frames in front of whoever generates adjacent shots. That reference discipline keeps the sequence in one coherent world.

Treat sound and edit as creative departments, not export settings. The editor sets rhythm and uses shot order to solve visual discontinuities; the sound engineer supplies impact and timing. AI provides more visual options. A three-minute live-show segment still needs people running it.

Did 86,000 people watch Kuzmin’s AI video?

Kuzmin said 86,000 people watched simultaneously at the first show. Treat that as his reported audience claim, not an independently verified count. The available account places the work on screens at two July 2026 Luzhniki shows and identifies Kuzmin’s contribution as one segment of the wider program.

The durable conclusion concerns attention design, not one attendance number. A between-songs film has a tight window to register with a crowd that did not come to inspect an advertisement or a film-festival entry. That brief rewards visual clarity, emotional scale, and edits that hit the music. It is demanding in any medium, AI video included.

For ecommerce and brand teams, look at the opening seconds of a product reel or paid social placement. Attention runs out fast. The generated image has to earn it with a stable product, readable action, and an obvious visual premise before decorative motion takes over.

What is the practical lesson for AI-video teams?

Commission AI video as a directed production with a large candidate pool; do not buy one generated clip and pray it carries the campaign. Kuzmin’s stadium segment used fifty-five 4K generations, four evenings of work, editorial assembly, and sound mixing to create roughly three minutes of screen time. That is serious work, and a highly flexible route to visual worlds at scale.

Keep people on the work that changes quality: setting the visual thesis, writing shot-specific prompts, maintaining continuity, rejecting weak candidates, cutting to music, and approving final brand-critical frames. A weak brief produces weak options. A precise one turns generation into a fast way to explore scenes, styling, and motion that would be difficult to stage conventionally.

Kuzmin’s move from architecture into AI video points to a wider creative advantage: people trained in composition, spatial thinking, and visual storytelling can carry those instincts into moving-image work quickly. The tool produces options. Taste decides what deserves the screen.