About

I am Xiang Wen. I am working toward real-time interactive virtual worlds: environments generated as you move through them, that change because of what you did.

Almost everything a generative model produces today is finished before anyone sees it. You write a prompt, you wait, you watch the result. The version worth building is the one where the world is still being produced while you are inside it. That needs a model holding a consistent world rather than a consistent clip, and it needs that model to answer fast enough to feel like a response. Neither is solved.

The problem does not belong to one field. It needs generation quality from generative AI and diffusion models, spatial consistency and real-time rendering from computer graphics, and geometry and camera understanding from 3D vision. I have shipped work in all four over the past decade, and my honest read is that the hard parts do not sit inside any of them. They sit in the seams, where a diffusion model has to respect geometry it was never taught, or a renderer has to accept something a network hallucinated.

Three generated frames: a robot dog on a pedestal, a teddy bear under a spotlight, and an apple lit from above — Frames from my paper-to-video pipeline. Every one of them was complete before anyone saw it. That property is the thing I want to remove.

How I got here

I started at NetEase Fuxi in 2016 on game agents, making an AI fight like a person rather than a script that chases you. We built a hierarchical reinforcement learning framework that split "when to engage and with whom" from "how to fight". It shipped, and it taught me something I have never forgotten: get the reward function slightly wrong and the agent will find a way through it that you never imagined.

In the same period I worked on music-to-dance generation, AI motion capture, and expression retargeting. None of these has a unique correct answer. The same piece of music supports a hundred reasonable dances, and no loss function will tell you which one is right. That is where I first understood that generation and discrimination need fundamentally different evaluation.

At ByteDance from 2020 the question became production scale. A large game ships an enormous amount of narrative animation every year while artist capacity grows linearly. We reframed character customization as a search problem in latent space, and used semantic graph matching to assemble narrative animation automatically. The lasting lesson was that whether a model can enter a production pipeline depends on how it fails, not on how it succeeds. A model with a high average and occasional catastrophes is unusable. A model with an ordinary average and a stable floor will ship.

At SkyReels AI in 2023 I ran a business rather than an algorithm team for the first time, as its business lead and CTO. The product asked whether someone who knows nothing about filmmaking could still turn a story into a short drama. Script generation, storyboarding, shot planning, video generation: four stages, each needing its own model, protocol, and evaluation. It was also where I first ran a full video post-training loop end to end, and first saw the uncomfortable gap between real user preference and offline metrics.

Speaking at the Apsara Conference in 2024 in front of slides on AI workflows, script AI, and storyboarding — Apsara Conference, 2024. The three questions on the slides behind me are still the ones I would put up today.

What I am working on

At Tencent Hunyuan I lead post-training for anime-domain video generation. Three threads: generation quality, a unified architecture handling text, image, multi-reference, and editing in one model, and inference acceleration.

That last thread is not a footnote. Interactive content is gated on latency before it is gated on quality. A model that produces a beautiful clip in ninety seconds cannot be stepped into. Most of the distance between where video generation sits now and where interactive content needs it is measured in milliseconds, not in benchmark points.

The open question I keep returning to is where the evaluation signal comes from. Nearly ten years separate the reward functions I hand-wrote for game agents from the preference models I work with now, and the underlying problem has barely moved:when "good" cannot be defined precisely, how do you construct a signal honest enough to steer a very large model. Interaction makes it harder, because what gets judged is no longer an artifact but a response.

On the side I am building a pipeline that turns papers into narrated videos. The reason is simple. The hard part of reading a paper was never the math, it was building intuition for the method as a whole, and video does that far better than text. The frames above came out of it.

If you are working on any of this, or think I have the wrong end of it, I would like to hear from you. Contact.