Research code that streams video generation chunk by chunk from a rough 3D proxy, built on Code World Model.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Streaming Code World Model is a research implementation that turns a streaming version of Code World Model (CWM) into a causal video generator. Each 0.71 s chunk is generated from the start image, the proxy (a coarse 3D scene rendering) so far, and earlier chunks, with no access to future frames. It uses a sliding 6-chunk attention span with KV cache reuse via RoPE rotation.
Training uses diffusion forcing with a noisy history to slow long-horizon drift, followed by 4-step DMD self-forcing distillation. The README reports a 4-step model rendering a chunk in 5.1 s on one NVIDIA H200 at 1344×768, about 9× faster than 35-step sampling, with lower pixel MSE on four held-out CARLA clips. It is not yet real time. The training data are 342 CARLA clips (3.7 h).
Checkpoints are LoRAs for the released MiniMax-H3 model, hosted on Hugging Face, with sampling and training scripts in the repository. Inference needs an H200-class GPU (123 GB peak in bf16, 108 GB with FP8); 80 GB GPUs are not supported yet. The Slurm scripts assume 8-GPU nodes. The repository license was not stated in the material provided.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
World model that transforms a rough 3D environment into realistic video one segment at a time, producing a segment in 5.1 seconds on one H200. More efficient than Nvidia's earlier method; code and training scripts released.
Closest neighbours by meaning, not just by category.
Research
Research framework where a coding agent maintains world state as code, rendered to video by a fine-tuned MiniMax-H3 model.
Model
Open framework and dataset for turning video diffusion models into real-time, long-horizon interactive world models.
Model
Real-time interactive world model with compressed memory and hybrid action and semantic control, built on Wan 2.2.
Model · Tencent ARC Lab
Video world model from Tencent ARC Lab that turns an image or text into camera-controlled explorable scenes with 3D-aware memory.
Model · Runway
Runway's real-time interactive world model generating 720p/24fps video with audio, steered by text actions and camera input.
Model · NVIDIA
NVIDIA research system where a code game engine drives a two-step streaming MiniMax-H3 video model for real-time rendering.
