Giga-World-1

This repository contains checkpoints for GigaWorld-1, a world model specially optimized for robot policy evaluation, introduced in the paper GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation.

GigaWorld-1 provides open-source training, inference, data processing, checkpoint conversion, and LoRA merge workflows for robot world models.

• Paper: GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation
• Project Page: Project Website
• Repository: GitHub

Directory Structure

model/
├── README.md
├── assets/
│   └── main_page.png
├── before_stage1/                       # Base checkpoints before stage-1 training
│   ├── Wan2p1_1p3B-FunContro-GigaRobo-alpha-diffusers/
│   ├── Wan2p1_1p3B-FunControl-diffusers/
│   └── Wan2p2_5B-FunControl-diffusers/
├── stage1/                              # Stage-1 fine-tuned checkpoints
│   ├── nano/                            # small (1.3B) variant
│   └── pro/                             # large (5B)   variant
└── stage2_distill/                      # Stage-2 distilled checkpoints
    └── ...

Training pipeline overview

Model Overview

Stage-1 checkpoint structure

Each Stage-1 variant (nano / pro) contains two released artifacts: a full Diffusers-format checkpoint and a scene LoRA checkpoint.

stage1/{nano,pro}/
├── Giga-World-1-*-stage1_final-diffusers/         # full Diffusers checkpoint
│   ├── model_index.json                           # Diffusers pipeline index
│   ├── transformer/                               # DiT / video transformer weights
│   ├── vae/                                       # VAE weights
│   ├── text_encoder/                              # text encoder weights
│   ├── tokenizer/                                 # tokenizer files
│   ├── scheduler/                                 # scheduler config
│   ├── image_encoder/                             # image encoder weights
│   └── image_processor/                           # image preprocessing config
└── Giga-World-1-*-stage1_scene_lora/              # scene LoRA checkpoint
    ├── pytorch_lora_weights.safetensors           # LoRA weights for inference
    ├── transformer_full/                          # full transformer export
    ├── transformer_partial.pth                    # partial transformer checkpoint
    ├── pytorch_model/                             # training checkpoint shards
    ├── distributed_checkpoint/                    # distributed training checkpoint
    ├── scheduler.bin                              # training scheduler state
    ├── latest                                     # latest checkpoint pointer
    ├── zero_to_fp32.py                            # ZeRO checkpoint conversion script
    └── random_states_*.pkl                        # training random states

Quick Start

Hugging Face Repository

https://huggingface.co/GigaAI-Research/Giga-World-1

SDK Download

# Install Hugging Face Hub
pip install huggingface_hub

# Download the model snapshot via Hugging Face Hub
from huggingface_hub import snapshot_download

model_dir = snapshot_download(repo_id='GigaAI-Research/Giga-World-1')

Git Download

git lfs install
git clone https://huggingface.co/GigaAI-Research/Giga-World-1

Acknowledgements

We sincerely thank the open-source community and the projects that make this work possible.

Thanks also to many other open-source contributors for their tools, models, and community support.

Citation

Please cite the paper if you find it helpful:

@article{gigaworld2025,
  title   = {GigaWorld-1: A Roadmap to World Models for Robot Policy Evaluation},
  author  = {{GigaAI}},
  journal = {arXiv preprint},
  year    = {2025},
  eprint  = {2607.02642},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

License

This model is released under the Apache License 2.0 unless otherwise specified.