Hyunwook Choi

MS. PhD. Integrated Student @kaist_ai, working on World Models
RS Intern NAVER AI Lab

[Hyunwook Choi @HyunwookEE]

Huge thanks to my amazing collaborators Seungryong Kim at @KAIST_AI and Dongyoon Han and Sangdoo Yun at @NAVER_AI_Lab

🎥 Project page: https://t.co/pFHjOVaW1n
📄 Paper: https://t.co/Fp2qdchz8b

[Hyunwook Choi @HyunwookEE]

Does every token need to be denoised until the final step? 🤔
No! We introduce ⚡Speculative Forcing⚡, which accepts tokens as soon as they converge and keeps denoising only the rest

Project page: https://t.co/pFHjOVaW1n 🧵 (1/n)

[Hyunwook Choi @HyunwookEE]

World models

It works the same way. As the camera moves, regions the model has already seen converge quickly, so only the newly revealed parts need more steps. 2.55 steps per token on average, 1.35× faster, with camera accuracy on par with 4-step.

5/n

[Hyunwook Choi @HyunwookEE]

Long video generation

It holds up over 30 seconds. 1.72 steps per token on average, 1.54× faster, and drift over the rollout is lower than the 4-step baseline, not higher.

4/n

[Hyunwook Choi @HyunwookEE]

Short video generation

Most tokens converge within one or two steps, and the remaining steps go where they are still needed. 1.68 steps per token on average instead of 4, 1.58× faster.

3/n

[Hyunwook Choi @HyunwookEE]

How it works. Each intermediate estimate is a draft. A verifier predicts how far a token’s draft is from the full-step output. Within tolerance, the token is accepted and committed. Only the rejected tokens enter the next pass.

2/n

[Hyunwook Choi @HyunwookEE]

TLDR; How can a world model continuously observe regions beyond the actor’s current view? Placing panoramic observers where we want, it evolves and controls the actor’s unseen regions.

[Junyoung Seo @jyseo_cv]

World models have one eye: the camera.

They can evolve what they see, but the rest of the world quickly goes unobserved.

So what if we gave them more eyes, like LoL wards or StarCraft Observers, and let them move around the world?

We introduce World Observer!

TL;DR: Keep the off-screen world evolving. Joint actor-observer generation with freely placed and movable panoramic observers.

Proj page: https://t.co/mDFlz5gvKb

@KAIST_AI @hyunwookee

[Junyoung Seo @jyseo_cv]

World models have one eye: the camera.

They can evolve what they see, but the rest of the world quickly goes unobserved.

So what if we gave them more eyes, like LoL wards or StarCraft Observers, and let them move around the world?

We introduce World Observer!

TL;DR: Keep the off-screen world evolving. Joint actor-observer generation with freely placed and movable panoramic observers.

Proj page: https://t.co/mDFlz5gvKb

@KAIST_AI @hyunwookee

[Hyunwook Choi @HyunwookEE]

𝐒𝐞𝐨𝐮𝐥 𝐖𝐨𝐫𝐥𝐝 𝐌𝐨𝐝𝐞𝐥 has been selected as a 𝐒𝐩𝐨𝐭𝐥𝐢𝐠𝐡𝐭 𝐎𝐫𝐚𝐥 𝐏𝐫𝐞𝐬𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 at #ECCV2026 🔥
Feel free to drop by Room Palissad on Friday morning. I’d love to chat, especially about World Models! Feel free to DM me :)

[Junyoung Seo @jyseo_cv]

What if a world model could render not an imagined place, but the actual city?

We introduce Seoul World Model, the first world simulation model grounded in a real-world metropolis.

TL;DR: We made a world model RAG over millions of street-views.

proj: https://t.co/Bx4KUAqrRs