moworld

Checkpoint layout

.
├── config.json
├── run_inference.py
├── high_noise_model.safetensors
├── low_noise_model.safetensors
├── high_noise_model/
│   └── config.json
├── low_noise_model/
│   └── config.json
├── configuration.json
├── Wan2.1_VAE.pth
├── models_t5_umt5-xxl-enc-bf16.pth
└── google/
    └── umt5-xxl/
        ├── special_tokens_map.json
        ├── spiece.model
        ├── tokenizer.json
        └── tokenizer_config.json

config.json is the checkpoint manifest used by run_inference.py: it maps each expert to its weight file and configuration. The launcher reads this manifest, validates the paths and configuration JSON, and sets TRAIN_HIGH_MODEL_PATH and TRAIN_LOW_MODEL_PATH before calling the MoWorld source launcher. It does not change the underlying architecture configuration lookup.

The two expert weights are at the repository root. Their configuration files remain in the corresponding expert directories.

⚙️ Quick Start

Prepare a Linux environment with Python 3.10+, PyTorch 2.7.1, and compatible CANN, torch_npu, and TorchAir packages; see the setup guide.

Code availability: As of September 21, 2026, the public GitHub repository contains only a README. The installation, inference, and training commands below require the full MoWorld source tree and its submodules. Model weights can be downloaded independently. The linked GitCode guides could not be verified during preparation of this model card.

Installation

Once the full source is available, clone the repository:

git clone --recursive https://github.com/Moxin-Tech/moworld1.0.git
cd moworld1.0
export MOWORLD_ROOT="$PWD"

Install dependencies in your Ascend environment:

python -m pip install -e ./MindSpeed
git clone --branch core_v0.12.1 https://github.com/NVIDIA/Megatron-LM.git ../Megatron-LM
export PYTHONPATH="$(cd ../Megatron-LM && pwd):${PYTHONPATH:-}"
python -m pip install -e "./moworld[realtime]"

Download the model with the Hugging Face CLI:

python -m pip install -U huggingface_hub
export MOWORLD_WEIGHTS="$MOWORLD_ROOT/models/moworld"
hf download moworld1/moworld_1.0_checkpoint --local-dir "$MOWORLD_WEIGHTS"

To download the weights before the source code is available, choose an explicit local directory:

hf download moworld1/moworld_1.0_checkpoint --local-dir ./models/moworld

Prepare the base weights, T5, VAE, and tokenizer under models/base/ following the asset layout. This checkpoint repository includes models_t5_umt5-xxl-enc-bf16.pth, Wan2.1_VAE.pth, and google/umt5-xxl/; downloading them into MOWORLD_WEIGHTS does not automatically populate models/base/. Arrange these assets and any additional base checkpoints according to the source code's asset guide.

Run the commands below from moworld/:

export MODEL_PATH="$MOWORLD_ROOT/models/base"
export MINDSPEED_PATH="$MOWORLD_ROOT/MindSpeed"
cd "$MOWORLD_ROOT/moworld"

Inference

Run camera-controlled autoregressive generation on 16 NPUs with the matching four-step MoWorld weights:

IMAGE_PATH="examples/moworld/realtime_assets/test.png" \
SINGLE_PROMPT="A tranquil landscape viewed by a camera moving slowly forward." \
CAMERA_MOTION="In-3,Left-2,Right-2" \
OUTPUT_PATH="$MOWORLD_ROOT/outputs/realtime" \
python "$MOWORLD_WEIGHTS/run_inference.py" --moworld-root "$MOWORLD_ROOT"

The default launcher uses 15 DiT processes and one VAE worker. For parallelism settings, bidirectional inference, and DCP inference, see the inference guide.

The configuration paths are $MOWORLD_WEIGHTS/high_noise_model/config.json and $MOWORLD_WEIGHTS/low_noise_model/config.json. If your source version resolves a configuration relative to its weight file, update that lookup to these paths. End-to-end inference with this repository layout has not yet been verified.

Training

Prepare your feature data and examples/moworld/local_data/train_data.json using the data preparation guide. Convert compatible base checkpoints, then train the high- and low-noise experts:

bash examples/moworld/convert_weights.sh

export MM_DATA="$PWD/examples/moworld/local_data/train_data.json"
export TRAIN_ITERS=10 SAVE_INTERVAL=10
bash examples/moworld/pretrain_high.sh
bash examples/moworld/pretrain_low.sh

These are short pretraining runs on 16 NPUs. Checkpoints are saved under checkpoints/train/. See the training guide for configuration and the export instructions for safetensors conversion.

Deployment

An HTTP API or web UI requires a separate service layer. See deployment details.

Validate the downloaded checkpoint layout

python "$MOWORLD_WEIGHTS/run_inference.py" --check

This checks file paths and JSON syntax without loading weights or starting NPU inference. Download the full repository, including config.json and run_inference.py, using the command above.