Practical overview
When it fits
Use World R1 when researching or adapting a Python-based text-to-video training workflow that incorporates camera trajectories and 3D-aware rewards, and when the required GPU, CUDA, checkpoint, and reward-service setup is available.
Who it is for
- Researchers investigating text-to-video generation and geometric consistency.
- Developers studying reinforcement-learning fine-tuning, camera-conditioned generation, or 3D-aware reward methods.
Documented uses
- Fine-tune a text-to-video model using camera and reconstruction feedback.
- Study how camera instructions can guide initial video latents without changing the base model architecture.
- Evaluate a training approach that combines 3D-aware rewards with a general visual-quality reward.
- Use dynamic-only training phases to address motion diversity during 3D-consistency training.
Requirements & limits
The public GitHub repository identifies the project as MIT-licensed and provides its source code and README instructions.
- The documented setup requires Python 3.10+, CUDA, and a PyTorch build matched to the system driver.
- Training examples use a Wan2.1-T2V-14B-Diffusers checkpoint and GPU process configuration.
- The README lists separate core and 3D-reward dependency groups, with optional acceleration packages.
- The repository describes code and training workflows rather than a hosted generation interface.
- Reward servers may be launched locally for single-node training or provided as URLs when already running.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the World R1 research environment
Prepare the documented Python environment and install the repository's core and 3D reward dependencies.
Before you start
- Linux research environment with Python 3.10 or later
- CUDA and a PyTorch build matched to the driver
- Conda and access to install Python packages
Steps
Create and activate a clean Python 3.10 environment, then upgrade pip.
Install PyTorch and torchvision using a wheel index appropriate to the CUDA version.
Install the repository in editable mode and the listed core runtime packages.
Install the listed 3D reward and visualization packages; install xformers and bitsandbytes only if needed for the machine and base-model setup.
Expected outcome
The README's Python environment and named package groups are installed; this setup alone does not launch training.
Guide limitations
- The CUDA-compatible PyTorch wheel index must be selected for the machine.
- WAN-based utilities require a model checkpoint, which can be passed with --model or configured with WORLD_R1_WAN_MODEL.
- The README does not establish compatibility for a particular hardware configuration.
Launch single-node training with local reward servers
Use the documented launcher to start single-node training with local 3D and general reward services.
Before you start
- The repository environment and dependencies have been set up.
- A Wan2.1-T2V-14B-Diffusers checkpoint is available by local path.
- The machine has the CUDA devices and process capacity selected in the launcher settings.
Steps
Set MODEL_PATH to the checkpoint path.
Set SERVER_VISIBLE_DEVICES, TRAIN_VISIBLE_DEVICES, and NUM_PROCESSES for the available devices.
Run bash scripts/run_single_node.sh from the repository.
Expected outcome
The repository's single-node training workflow is launched with local reward-server configuration.
Guide limitations
- The README provides an example using separate device assignments and six training processes; adapt the settings to the machine.
- The guide does not establish successful execution or a resulting model artifact.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- microsoft/World-R1 · Official homepage title · retained in its source language