Studio Realive

Studio: Studio Realive by SM Entertainment
Role: Immersive Unit Lead (Jun 2026 – present) · Principal Manager, Content Producing Unit (Mar 2025 – May 2026)
Location: Seoul, S.Korea
Focus: Immersive experience & AI-driven IP universe expansion · AI production
Real-time: Unreal Engine 5 · Unity · custom multi-sensor interaction sw
Gen AI: Veo 3 (Flow) · Nano Banana · Kling · Seedance · Seedream · Midjourney · Higgsfield · ComfyUI · ElevenLabs
LLM: Claude · OpenAI · Gemini · Grok
Pipeline / tooling: Node + TypeScript engine · React + React Flow (Maker Console) · WebSocket · Strudel · three.js · Cloudflare R2 / Workers
Dev tools: Claude Code · Codex · Gemini CLI
Core languages: C++ · JavaScript · Python

I lead the Immersive Unit, setting creative direction and running production and technical validation. We design, build and pitch immersive experiences that extend SM universes across digital and physical spaces, and put generative AI into production. I am also working with other SM units, including the concert and video units, to discuss where we can use AI in their production pipelines to save time and money while keeping creative and visual quality.

Below are some examples of how we used AI in production. Most projects have not been announced yet, and are still under heavy iteration, continuously being improved.

Case study 01 · July 2025 · AI film for a media façade
Taehwagang, Ulsan: making fifteen minutes of AI film for a whale-shaped skywalk

Commissioned by Ulsan City · 2025 · Studio Realive · Delivered Sep 2025, installed Dec 2025. I was in charge of making the video.

5 pieces, ≈3 min each · 3000×3840 projection face · 9498×192 railing LED strip

The surface

This was a media façade on the Yonggeumso Skywalk at Taehwaru, with a tall 3000×3840 projection face and two LED railing strips at 9498×192 and 3166×64. I was making video for a structure people walk on at night, rather than a screen.

The film

There were six pieces of about three minutes each: a whale, the river spirit as a blue dragon, a yellow dragon, the four seasons in two versions, and the industrial city. Five were used in the installed show, fifteen minutes in all. Every frame was generated, then corrected by hand where the model's motion drifted.

The model at the time

Video generation models were improving fast, but they were still not good enough to make a continuous, single-shot film. The client wanted limited cuts and a continuous storyline focused on the elements they wanted to show. So we had to rely on Kling v1.6, which accepted start and end images as inputs. We could stitch the clips together without much trouble, but the objects kept moving at different speeds from clip to clip. This was very hard to control with prompts alone. I tested many prompts to keep objects moving at a consistent pace, and we also tried different start and end images to get the best results.

The workflow and corrections

Then came the part people don't talk about much: fixing keyframes where motion stuttered and rebuilding transitions where clips didn't meet.

The projection surface needed its own corrections. We shortened the white-out and black-out timing because a bright projected frame reveals the building behind it. We increased the contrast so the waterfall would be visible against the sky from the riverbank, and upscaled everything to the façade's native resolution. We worked clip by clip, with a written correction list for each piece.

Before / after

Fifteen minutes of projection content normally takes months of CG work. Here, generation was 95% done within about six weeks of the plan, and the remaining time went to making the content work on the surface.

Case study 02 · Sep 2025 · AI music video
aespa 'Rich Man (Yellow Claw Remix)': making an AI music video in two weeks, with no shoot

Released Sep 19, 2025 on the official SMTOWN channel · SM Entertainment × Studio Realive, in official collaboration with YouTube and Google DeepMind · Made in Flow on Veo 3.0 · No camera, no set, no cast. Team: creatives and artists working alongside video editors. My part: AI video production and pipeline research.

0 cameras, sets, cast · 2 wk storyboard → delivered master · 3 phases: test · board · generate+cut

1 · Testing what Veo 3.0 could actually do

Before we wrote a storyboard, we tested Veo 3.0, which was unreleased at the time, across every style we could think of, including realistic, animated and abstract. Veo 3.0 was state of the art at the time, but it still had limits. We wanted to understand its actual capabilities beyond the demo reel. For a music video, two limits mattered: believable human movement and control over what happens inside a shot.

2 · Writing a storyboard that met the brief and the model could finish

Our first instinct was to go psychedelic, using an animatic style where the model's weaknesses could become part of the look. The direction from SM and Google was the opposite: realistic people dancing on screen.

We resolved this through the story. The remix is by Yellow Claw, so we made a family film about the DJ. Each family member dances to the track while going through their own day, and everyone reunites at dinner. Short, self-contained dance beats with one subject per shot play to what the model does well, while avoiding long choreography and precise camera moves. The storyboard was deliberately “safe.” Every shot was something we already knew Veo could deliver, and we scoped the whole thing to finish in two weeks. In hindsight, I would have pushed for the animatic direction, since the audience response suggested the realistic look wasn't what fans wanted.

Before we started generating, we made shared assets: character sheets, style reference images and prompts describing each person and their dance moves. We all used these assets to generate our clips, keeping the characters and style consistent across work made by different people. Everything felt like part of the same music video.

3 · Generating and editing at the same time

We started clip generation and editing together. The editors cut with the selected clips while new shots were still being generated, so the timeline shaped the prompts as much as the prompts shaped the timeline.

Before / after

A conventional MV of this shape would usually need a cast, locations, a shoot day per setting, then post. A full-CG version would have used simple 3D model animation with some VFX, or it could have taken months to create the assets. We had no shoot and no 3D at all. A team of creatives, artists and editors took it from storyboard to delivered master in a two-week production window.

What I would tell another production

Test the model before writing for it. Make creative decisions around what the tool does well, and let the story absorb the rest. Size the storyboard to what you have actually seen the model finish, rather than what its best sample suggests. And start the edit on day one.

Case study 03 · Sep 2026 · AI installation
Naevis 1998: merging AI model into coworking canvas

EnterTech Seoul 2026 · Sept 18–20 · DDP · Studio Realive · A graphic chatbot I designed and built in six weeks with Claude Code. Creative content by the Immersive Unit's creative team.

212 sessions in 3 days · 84% finished the chat · 147 pieces, all four assets · 0 publication failures · $0.34 per visitor session

The experience

The brief was a convention installation that was not obviously a chatbot. No keyboards, no personal data and no artist impersonation. A visitor sits at an iPad, and a portrait TV shows the world of a hidden entity, Naevis, who notices them first. Naevis and a virtual entity from 1998 communicate with the visitor. They write a short poem together, Naevis gives it a title, and the visitor gets a cover image, a 40-second song and a 5-second video on their phone via QR. The experience takes three to five minutes, then the kiosk resets itself.

We wanted to explore using multiple generative AI models in one experience with strong visual feedback. The idea was a graphic chatbot, like a graphic novel.

The workflow

I connected seven models in one graph. A fast model handles the chat turns, a second writes the poem, and a third writes the title. Image, music and video generation then run in parallel. There is also live-coded sound during the chat. The sound is procedural on purpose, so the interactive part never waits on a network.

Everything critical runs on the show machine, and every cloud call has a rehearsed local fallback. My rule from day one was that the show should never stall in front of a visitor. The interaction should feel like talking to an actual person rather than waiting for a generated response delayed by the network.

The tool and the handoff.

The creative team needed to change Naevis's personality, the poem rules, the pacing and the sound presets without an engineer, right up to and during the show. So I made the generation pipeline a node graph, with the Maker Console as its editor. I had used ComfyUI before, and I built the console as an easier-to-use version of it for the team, with a shared canvas, shared assets and saved nodes.

The console also makes it easy to save frequently used node chains as reusable functions. An asset library lets us easily reload generated outputs. The functions defined in the graph tool were later used in the actual chatbot.

Every model is a typed node, and every intermediate result can be inspected. There is a dry-run bench, and edits hot-reload into the next session. From the fourth week, the creative team handled content iteration, and technical work was limited to shipping fixes.

Before the build

Every AI behavior assumed in the concept deck was tested against the real APIs before production started. Each was written up as works / doesn't / needs a rule. Three assumptions didn't survive contact with the APIs. Licensing was not an issue for sharing the generated content with visitors. The risk was generation corrupting the company's IP, so we added an IP-protection AI step.

Results

The kiosk ran 212 sessions over three days. 84% of visitors who started the chat finished it, and there were 147 finished pieces with all four assets. It took about five minutes per visitor, with 8 to 12 sessions an hour through the middle of each day.

The median time from the end of the poem to the last asset on the visitor's phone was 92 seconds. There were zero publication failures. On the show days, the cost was $0.34 per visitor session across four model providers. A per-provider ledger also showed that the largest line item was one nobody had metered.

One takeaway was that people lost interest as soon as they saw a traditional chat window. We are planning to create an animated character that poses based on AI's real-time interpretation of the conversation.

Also at Studio Realive

Global LBE. I led concept and development through full technical validation for immersive experiences in Japan (a transforming space that replaces the movie theater), Saudi Arabia (projection, lighting and LED-wall performance content for a giant mall, plus its media-server design) and Vietnam (a new retail experience for a mall under construction). The first two were in the ₩10–50B project-cost range. I built multi-sensor interaction tools and Unreal content along the way.

Expanding an IP universe with AI. With a global AI partner, I broke aespa's Kwangya universe down into structured data covering places, rules, events and beings. I built a world-onboarding process on that data, then defined the universe's creative rules and a Creative Graph: the base for generation that stays inside the world. The Naevis persona and voice pipeline in the kiosk grew out of the earlier Naevis chatbot and TTS avatar work.