Rocket launch. Model output, first frame.

Structure from the proxy, appearance from the first frame

A proxy sequence carries the layout, object classes and motion of a scene, at any level of detail. A reference first frame supplies its appearance. Each chapter below changes one input and holds the others fixed. Every clip on this page is model output, shown with the proxy that drove it.

01

Camera control

Each pair opens on one first frame of one scene, then the proxy flies two different routes: a dive down a glass tower, a dolly zoom on a bridge, a chase out of a tunnel, a drone beside a rocket launch. The video takes whichever route its proxy flies.

Fixed: first frame, scene Varies: camera route

Tower dive. Proxies for routes A and B above, model outputs for each route below, first frame.
Tower dive Both routes open on one first frame, then split. Each proxy is inset at top left.

02

World dynamics

The camera holds its path and the proxy changes what happens in the world. At an intersection, the cross traffic moves on one signal plan and the pedestrians cross on the other. At a bascule bridge, the two leaves lift and a cargo ship passes through, or the bridge stays shut and the traffic keeps flowing. The video follows each version of the proxy.

Fixed: first frame, camera Varies: motion in the scene

Intersection. Proxies for versions A and B above, model outputs for each version below, first frame.
Intersection Both versions open on one first frame and share the camera. Only the motion in the proxy differs. Each proxy is inset at top left.

03

Proxy correspondence

Every element in the proxy lands at the same place in the video: the pillars, the temple and the figure through a dolly zoom, the colonnades and the statue of a sunken street as the camera rises to the sea surface, the parts of a motorcycle, the lamps of a car as the camera circles behind it. Fade the proxy over the output to compare them in place.

Fixed: one proxy and first frame per scene Varies: scene

Cliff temple. Proxy and model output, first frame.
Cliff temple Proxy drawn over the output at the chosen opacity, frame-synced.

04

Any granularity

One first frame and one camera route, with four proxies at different levels of detail. Each carries a different amount of scene structure into the video.

Fixed: first frame, camera route Varies: proxy granularity

Rainy chase. Proxy and model output, first frame.
Rainy chase Proxy inset at top left, frame-synced with the output. Switching proxy keeps the playback position.

05

Game demos

The same model renders game footage from game-engine proxies: a getaway down a coastal strip that drifts into a parking garage, a supercar dropped from a cargo plane onto a desert highway, a jet threading a red-rock canyon, an anti-gravity racer through a corkscrew and up a spiral ramp. Two bullet-time shots stop the world and move only the camera: a car bursting through a showroom window, and a city folding shut over the street.

Fixed: model, proxy style Varies: game, scene, motion

Coast getaway. Proxy and model output, first frame.
Coast getaway Proxy inset at top left, frame-synced with the output.

06

Play it live

Racing session. Model output full frame, proxy at bottom left, the player's controller at bottom right.

Infra

Real-time streaming

The video is generated chunk by chunk, and the final pipeline delivers a chunk every 930 ms, 18.0 frames per second, 1.82× the 1770 ms baseline. Each chunk takes four denoising passes of the DiT on 8 GPUs (Ulysses8). The first pass is fused with the commit of clean history and proxy condition into the KV cache, which keeps a sink and a recent window and evicts the rest. Between DiT chunks, a lightweight TAE encodes the next proxy chunk while the decoder turns the current one into RGB, and a CPU worker delivers those frames while the DiT renders the next chunk.

8 GPUs · Ulysses8 Fused pass 1 + KV commit Concurrent encode / decode Async delivery

Infra overview. A KV cache holds a sink, evicted slots, earlier chunks of clean history H and proxy condition C, and the newly committed KV. Two DiT blocks, each four passes on 8 GPUs with Ulysses8, read the cache; pass 1 is fused with the commit of H and C. Between them, a codec interval runs the encoder (×2) and decoder (×4) concurrently, taking proxy chunk n+1 in and putting RGB chunk n out.
Infra overview. Concept diagram, not drawn to time scale. H clean history · C proxy condition · D noisy latent.
Stage timeline of the final pipeline over three consecutive chunks, 0 to 3 seconds. Lanes: proxy encode with TAE on 1 GPU, DiT render with Ulysses-8, decode and device-to-host copy with causal TP4, CPU delivery on an async worker, and frames out. Each chunk runs a fused pass with clean history, proxy condition and the first noisy latent, then three more denoising passes. Between chunks, encoding of the next proxy chunk overlaps decoding of the current one, and CPU delivery of a chunk overlaps the DiT render of the next. Frames out: 21, 13 and 21 frames, about one chunk period apart.
Stage timeline of the final pipeline over three consecutive chunks, laid out from a profiled run on one time axis. Dashed boxes mark work that runs concurrently.

Citation

@misc{wang2026codeworldmodel20,
  title         = {Code World Model 2.0: Towards Infinite Controllable Video Generation},
  author        = {Ruoyu Wang and Hao Zhong and Yanming Yang and Jiahao Tian and Lin Liu and Biao Gong and Guoshan Lu and Yiding Tian and Chunhua Shen and Zhenzhong Lan and Chi Zhang},
  year          = {2026},
  eprint        = {2611.22925},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2611.22925}
}