Moxin-Tech / moworld_base

MoWorld: A Flash World Model 
Team Moxin
https://github.com/user-attachments/assets/57cfb64f-4103-4070-834c-1801c277e329
---
We present **MoWorld**, a cost-efficient, high-performance Flash World Model for real-time interactive video generation on Neural Processing Units (NPUs). MoWorld jointly optimizes data, algorithms, systems, and hardware across an end-to-end pipeline spanning data generation and pretraining.
- **3D-Native Data Engine**: A scalable data pipeline constructs geometrically consistent training samples from real-world and synthetic environments through geometry completion, quality control, trajectory verification, and vision-language annotation.
- **Curriculum Cross-Frame Pretraining**: Training progressively extends from short clips to sequences of up to 2,000 frames, improving camera control, temporal consistency, and long-horizon spatial memory.
- **Few-Step Autoregressive Generation**: Autoregressive flow-matching pretraining and Self-Forcing distillation reduce the standard 50-step sampling process to 4-step generation.
- **Real-Time NPU Inference**: Pipeline-, parallelism-, and kernel-level optimizations enable up to 50 FPS on NPUs in the configuration reported in the paper, with an average inference cost reported at 30%–50% of existing world-model solutions.
Given a first frame, a text prompt, and camera controls, it generates long-horizon video while preserving visual quality, spatial consistency, and control responsiveness. Visit the project page to explore more world-generation, reconstruction, and downstream application results.
## 🔥 News
- Sep., 2026: We open source MoWorld and will continue to release subsequent versions of MoWorld.
- Jul. 17–20, 2026: MoWorld debuts and exhibits at the World Artificial Intelligence Conference (WAIC 2026) in Shanghai.
- Jul. 7, 2026: We released the first version of the MoWorld technical report at arXiv.
## ⚙️ Quick Start
Prepare a Linux environment with Python 3.10+, PyTorch 2.7.1, and compatible CANN and `torch_npu` packages; see the setup guide.
### Installation
Clone the repository:
 [code] 
Install dependencies in your Ascend environment:
 [code] 
Download the MoWorld Checkpoint:
 [code] 
Prepare your training data and any required tokenizer/encoder assets following the asset layout, then run the commands below from `moworld/`:
 [code] 
### Training
Prepare your feature data and `examples/moworld/local_data/train_data.json` using the data preparation guide. The published checkpoint contains the high-noise expert only. Link it into the converter's expected layout, convert it to DCP, then train the high-noise expert:
 [code] 
This is a short pretraining run on 16 NPUs. Checkpoints are saved under `checkpoints/train/high_noise_model/`. See the training guide for configuration.
## 💡 Highlights
https://github.com/user-attachments/assets/be71859e-8075-4c04-bbab-068b33cb16a9
https://github.com/user-attachments/assets/de0d19e8-e07e-4f31-b80a-791b816132a3
https://github.com/user-attachme