open-gigaai/giga-world-1

What it solves

GigaWorld-1 provides a framework for building and training world models specifically designed for robot policy evaluation. It addresses the challenge of creating controllable, high-fidelity video simulations of robotic tasks, allowing researchers to test robot policies in a virtual environment before deploying them in the real world.

How it works

The project implements a multi-stage training pipeline based on the WAN architecture:

• Stage-1 (Controllable Pre-training): Trains the Nano (1.3B) and Pro (5B) models to generate controllable video sequences.
• Stage-2 (Acceleration Distillation): Uses DMD2 to distill the denoising process, reducing the required inference steps from 20 down to 4–6, which significantly speeds up video generation.
• Data Pipeline: Processes robot data (LeRobot-style) into a GigaWorld format using Qwen3-VL for captioning and Depth Anything V2 for depth estimation.
• Inference: Supports both image-to-video (i2v) and text-to-video (t2v) generation, capable of producing 33-second rollouts at 10 FPS.

Who it’s for

It is designed for robotics researchers and AI engineers who need high-quality world models to evaluate robot policies and those looking to train world models on custom robotic domains.

Highlights

• Dual Model Sizes: Offers both Nano (1.3B) and Pro (5B) versions to balance performance and resource constraints.
• Efficient Inference: Stage-2 distillation reduces generation steps to 4–6 for faster rollouts.
• Comprehensive Tooling: Includes open-source workflows for training, inference, data processing, and LoRA merging.
• Domain Adaptability: Allows adaptation to new robotic domains by simply providing data in the GigaWorld format without requiring code changes.