README.md

---
title: Streaming Code World Model
emoji: 🌍
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Causal chunk-by-chunk world-model video streaming
python_version: "3.12"
startup_duration_timeout: 1h
models:
  - Luffuly/streaming-code-world-model
  - Comfy-Org/MiniMax-H3
---

# Streaming Code World Model

Demo of [Streaming Code World Model](https://github.com/TingtingLiao/streaming-code-world-model)
([checkpoints](https://huggingface.co/Luffuly/streaming-code-world-model)): causal, chunk-by-chunk video generation
from a live control proxy, built on [Code World Model](https://github.com/buaacyw/code-world-model)
(MiniMax-H3 Ref2VA + CWM LoRA).

Sampling follows `rtcwm.eval.open_world generate --steps 4 --rule x0 --shift 5 --kv-reuse --fp8` with
`streaming_dmd_4step.safetensors`. Adaptations for ZeroGPU (RTX PRO 6000 Blackwell): FlashAttention-3 / flex-attention
replaced by PyTorch SDPA over the same block-causal key prefixes, FP8 block linears (the authors' `--fp8`), and the
per-block AdaLN projections precomputed for the sampler's fixed set of timesteps.

Scenes are released Code World Model inference examples
([NTU-yiwen/code-world-model-inference-examples-40](https://huggingface.co/datasets/NTU-yiwen/code-world-model-inference-examples-40)):
their start frame, the "Appearance and scene" caption of the first window and the Depth + Semantic-ID proxy, pre-encoded
(Qwen3-VL-32B states, VAE latents) exactly as `rtcwm.eval.open_world prepare` does.

Vendored code: `musubi_tuner/` (kohya-ss/musubi-tuner @ c4dc6b8 + the CWM `minimax-h3-retake34` and headless patches,
Apache-2.0), `cwm_h3_inference/` (buaacyw/code-world-model, Apache-2.0), `rtcwm/` (TingtingLiao/streaming-code-world-model).