---

📑 Table of Contents

• 📰 Latest Updates

• 📊 Open-Source Progress

• 1. 📦 Environment Setup

  • 1.1 🖥️ Hardware & OS
  • 1.2 🧰 Install Dependencies

• 2. 🗃️ Data Preparation

• 3. 🧩 Model Preparation

• 4. 🚂 Training

  • 4.1 Stage-1 Training
  • 4.2 Stage-2 DMD Training
  • 4.3 Training on Other Domains

• 5. 🎬 Inference

• 6. 🔄 Model Merge & Checkpoint Conversion

• 7. 📁 Repository Layout

• 🙏 Acknowledgements

• 📖 Citation

---

📰 Latest Updates

💡 Subscribe to releases — click Watch ▾ → Custom → Releases on the GitHub repo to be notified when new weights, datasets, or the WMBench benchmark drop.

---

📊 Open-Source Progress

🟢 Released · 🟡 Beta · 🔴 Coming Soon — last updated 2026-07

🌐 Release Channels

🛠️ Want a component to ship sooner? Please file an issue or open a PR.

---

1. 📦 Environment Setup

1.1 🖥️ Hardware & OS

Note: We use a single-node 8× H20 or 8× A100 setup for production training. With appropriate memory optimization techniques, the released code can also run training and inference experiments on consumer-grade GPUs.

1.2 🧰 Install Dependencies

Install PyTorch first according to your CUDA version, then install the remaining dependencies with install.sh.

cd <PROJECT_ROOT>

# CUDA 12.6
pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cu126

# CUDA 12.8
pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cu128

# CUDA 13.0
pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cu130

# Install dependencies
bash install.sh

install.sh installs packages from requirements.txt.

Main dependencies:

accelerate>=1.1.0        # accelerate launch / DDP / DeepSpeed
av>=12.0.0               # video read / write
decord>=0.6.0            # fast video decoding
diffusers>=0.35.0        # custom-modified version (see thirdparty/)
einops>=0.8.0
imageio>=2.36.0
imageio-ffmpeg>=0.5.1
numpy>=1.24,<3
omegaconf>=2.3.0
opencv-python>=4.9.0
packaging>=24.0
pandas>=2.1.0
peft>=0.12.0             # LoRA implementation
Pillow>=10.0.0
pyyaml>=6.0.0
safetensors>=0.4.5
torchdata>=0.8.0
tqdm>=4.66.0
transformers>=4.45.0
wandb>=0.18.0            # offline by default
xformers>=0.0.28.post3   # memory-efficient attention
deepspeed==0.19.2
ftfy==6.3.1

Optional: install FlashAttention 2 or 3

FlashAttention 2 or 3 can be used depending on your CUDA / PyTorch environment. Install one compatible version if you need FlashAttention acceleration.

# FlashAttention 2
pip install flash-attn --no-build-isolation

# Or build FlashAttention 3 manually
cd thirdparty/flash-attention-3
# Follow its README

---

2. 🗃️ Data Preparation

This release provides a small toy data package for verifying inference, data loading, visualization, and training workflows. The toy data is available from Hugging Face and ModelScope.

Use the one-click downloader:

bash tools/download_tool/download_giga_world.sh \
  --platform hf \
  --target toydata \
  --output-dir ./downloads

For ModelScope, replace --platform hf with --platform modelscope. See tools/download_tool/README.md for all options.

After downloading, place or symlink the toy data under example/:

mkdir -p example
cp -r ./downloads/Giga-World-1-Toydata/* ./example/

Expected project structure:

giga-world-release/
└── example/
    ├── toy_datapipeline_dataset/    # raw LeRobot-format toy dataset
    │   ├── gt/
    │   ├── depth/
    │   ├── plucker/
    │   ├── sketch/
    │   └── labels/
    └── toy_train_dataset/           # model training data
        ├── nano/
        │   ├── dataset_cache.pkl
        │   └── episode_*.pt
        └── pro/
            ├── dataset_cache.pkl
            └── episode_*.pt

toy_train_dataset/ is already in the format used by the training configs: stage_1_post_functrl_wan21.yaml, stage_1_post_functrl_wan22_5b.yaml, and stage_2_dmd_functrl_wan21.yaml.

For raw data visualization, run the web tool and open http://127.0.0.1:8090/ or http://127.0.0.1:8090/calib:

cd tools/data_vis_tools
python app.py --host 0.0.0.0 --port 8090

For offline latent pre-computation, use get_short-latents-giga-ctrl.py or get_short-latents-giga-ctrl-wan22-5b.py. The input data should contain helios_giga_ctrl.jsonl, videos/, and control_videos/; outputs are .pt samples containing precomputed vae_latent, control_latent, prompt_embed, and related metadata. See tools/offload_data/data_format.md for the data format.

3. 🧩 Model Preparation

Released model weights are available from:

Use the download helper:

• Script: download_giga_world.sh
• Full usage: tools/download_tool/README.md

Download model weights from Hugging Face:

bash tools/download_tool/download_giga_world.sh \
  --platform hf \
  --target model \
  --output-dir ./downloads

Download model weights from ModelScope:

bash tools/download_tool/download_giga_world.sh \
  --platform modelscope \
  --target model \
  --output-dir ./downloads

After downloading, place or symlink the model files under model/:

giga-world-release/
├── model/
│   ├── before_stage1/
│   │   ├── Wan2p1_1p3B-FunContro-GigaRobo-alpha-diffusers/
│   │   ├── Wan2p1_1p3B-FunControl-diffusers/
│   │   └── Wan2p2_5B-FunControl-diffusers/
│   ├── stage1/
│   │   ├── nano/
│   │   └── pro/
│   └── stage2_distill/          # coming soon
└── tools/
    └── download_tool/

If the downloader saves weights under ./downloads/Giga-World-1/, copy or symlink them into the repository:

mkdir -p model
cp -r ./downloads/Giga-World-1/* ./model/

---

4. 🚂 Training

flowchart LR
    %% Base Models
    A1["WAN 2.1<br/>1.3B FunControl"]
    A2["WAN 2.2<br/>5B FunControl"]

    %% Before Stage1
    B1["GigaRobo Alpha<br/>Diffusers"]
    B2["WAN2.1 1.3B<br/>Diffusers"]
    B3["WAN2.2 5B<br/>Diffusers"]

    %% Stage1
    C1["Nano"]
    C2["Pro"]

    %% Stage2
    D1["Nano Distill"]
    D2["Pro Distill"]

    %% Connections
    A1 -- "Pretrain on Giga Dataset<br/>Convert to Diffusers" --> B1
    A1 -- "Convert to Diffusers" --> B2

    B1 --> C1
    B2 --> C1
    C1 --> D1

    A2 -- "Convert to Diffusers" --> B3
    B3 --> C2
    C2 --> D2

    subgraph S0["Before Stage 1"]
        subgraph S0A["Interchangeable"]
            B1
            B2
        end
        B3
    end

    subgraph S1["Stage 1"]
        C1
        C2
    end

    subgraph S2["Stage 2"]
        D1
        D2
    end

    classDef base    fill:#4F8EF7,color:#fff,stroke:#2D5FD2,stroke-width:3px;
    classDef diff    fill:#7C4DFF,color:#fff,stroke:#5B2DCC,stroke-width:3px;
    classDef stage   fill:#00C853,color:#fff,stroke:#009624,stroke-width:3px;
    classDef distill fill:#FF6D00,color:#fff,stroke:#DD2C00,stroke-width:3px;

    style S0A fill:none,stroke:#FF0000,stroke-width:3px,stroke-dasharray:8 4,color:#FF0000;

    class A1,A2 base;
    class B1,B2,B3 diff;
    class C1,C2 stage;
    class D1,D2 distill;

Training entrypoints and launcher scripts are paired (each pair = one accelerate launch command + one YAML).

4.1 Stage-1 Training (Controllable Pre-training)

Launch Nano:

bash scripts/training/stage1/train_deepspeed_stage1_functrl_wan21.sh

Launch Pro:

bash scripts/training/stage1/train_deepspeed_stage1_functrl_wan22_5b.sh

Default output layout:

output/
├── exp/
│   ├── Giga-world-Nano-Train-Stage-1/
│   └── Giga-world-Pro-Train-Stage-1/
└── logs/

4.2 Stage-2 DMD Training (Acceleration Distillation)

DMD2 compresses the denoising loop from 20 steps to 4–6 steps (the Stage-2 config sets num_inference_steps: 6) and aligns with a frozen real score model via a critic_lora.

Launch Nano DMD:

bash scripts/training/stage2/train_deepspeed_stage2_functrl_wan21.sh

Example output:

output/exp/Giga-world-Nano-Train-DMD/

4.3 Training on Other Domains

To adapt GigaWorld-1 to a new domain, no code changes are required. In practice, you only need to construct training data in the same format as the provided GigaWorld data, then launch the existing Stage-1 training pipeline with the corresponding config.

For most new domains, training for a few thousand steps on 8 GPUs is typically sufficient and can usually finish within one day.

5. 🎬 Inference

Usage:

# Nano i2v: first frame + control video + text prompt
bash scripts/infer/run_infer_nano_i2v.sh

# Pro t2v: text prompt only (omit --image_path → t2v mode)
bash scripts/infer/run_infer_pro_t2v.sh

The underlying entrypoint infer_giga_world.py exposes the following arguments:

Inference output example:

---

6. 🔄 Model Merge & Checkpoint Conversion

Use uni_merge_lora_for_giga_world_1.py to merge LoRA or partial checkpoints into a standalone transformer for deployment.

The tool supports both wan2.1 and wan2.2_5b, automatically resolves checkpoint files from the specified checkpoint directory, and saves the merged transformer to --save_dir.

python tools/ckpt_tools/uni_merge_lora_for_giga_world_1.py \
  --base_model <PATH_TO_BASE_NANO> \
  --save_dir   <PATH_TO_STAGE1_MERGED_NANO> \
  --ckpt_dir   /path/to/checkpoint-XXXX \
  --model_type wan2.1

For Pro 5B:

python tools/ckpt_tools/uni_merge_lora_for_giga_world_1.py \
  --base_model <PATH_TO_BASE_PRO> \
  --save_dir   <PATH_TO_STAGE1_MERGED_PRO> \
  --ckpt_dir   /path/to/checkpoint-XXXX \
  --model_type wan2.2_5b

---

7. 📁 Repository Layout

.
├── gigaworld/                         # Core model, pipeline, data loader, scheduler, and utils
│   ├── dataset/                       #   Stage-1 / Stage-2 / DMD data loaders
│   ├── modules/                       #   Transformer + custom Triton / Flash kernels
│   │   ├── gigaworld_kernels/         #     fp32_rmsnorm, tiled_linear, triton_norm, triton_rope
│   │   ├── transformer_gigaworld.py
│   │   └── transformer_functrl_gigaworld.py
│   ├── pipelines/                     #   i2v / t2v main pipelines
│   ├── scheduler/                     #   custom schedulers
│   ├── utils/                         #   TrainConfig, EMA, recycle batch, etc.
│   └── videoalign/                    #   reward / VLM training & inference
├── infer/
│   └── infer_giga_world.py            # Python inference entrypoint
├── scripts/
│   ├── accelerate_configs/            # DeepSpeed ZeRO-2 / ZeRO-3 configs
│   ├── infer/                         #   i2v / t2v × Nano / Pro one-click scripts
│   ├── training/
│   │   ├── configs/                   #   Stage-1 / Stage-2 YAMLs
│   │   ├── stage1/                    #   Nano / Pro Stage-1 launchers
│   │   └── stage2/                    #   Nano / Pro Stage-2 DMD launchers
├── tools/
│   ├── ckpt_tools/                    #   LoRA merge and checkpoint utilities
│   ├── datapipeline/                  #   LeRobot-style data preprocessing
│   ├── data_vis_tools/                #   Web URDF + camera calibration viewer
│   ├── download_tool/                 #   one-click HF / ModelScope downloader
│   ├── offload_data/                  #   offline latent pre-computation / format conversion
│   └── others/                        #   misc conversion tools
├── assets/                            #   README figures and demo media
│   ├── main_page.png                  #   main teaser image
│   ├── data_vis.gif                   #   data visualization demo
│   ├── input_image.png                #   inference first frame
│   ├── control_video.mp4              #   inference control video
│   └── i2v_sample.mp4                 #   generated rollout sample
├── example/
│   ├── toy_train_dataset/             #   Nano / Pro toy training datasets
│   └── toy_datapipeline_dataset/      #   toy preprocessing output (gt / depth / plucker / sketch)
├── model/
│   ├── before_stage1/                 #   Diffusers-converted base checkpoints
│   ├── stage1/                        #   Nano / Pro Stage-1 checkpoints
│   └── stage2_distill/                #   distilled checkpoints (coming soon)
├── train_gigaworld_functrl_uni_stage1.py
├── train_gigaworld_functrl_uni_stage2_dmd.py
├── requirements.txt
└── install.sh

---

🙏 Acknowledgements

GigaWorld-1 stands on the shoulders of a vibrant open-source ecosystem. We are deeply grateful to the following communities and projects that made this work possible:

🤗 Foundation Models & Architectures

• Wan (Alibaba) — the wan2.1 and wan2.2_5b backbones that power GigaWorld-1 Nano and Pro
• Diffusers — the modular diffusion framework we extend with custom attention processors and pipelines
• Helios — a video generation model that achieves minute-scale, high-quality video synthesis
• Genesis — a generative universal physics engine and robotics/embodied AI simulation platform
• Hugging Face 🤗 — hosting, transformers, accelerate, and the entire model & dataset ecosystem
• open-gigaai/Giga-World-1 and open-gigaai/Giga-World-1-Toydata — the public Hugging Face model and toy-data repositories

🛠️ Training & Acceleration

• PyTorch & DeepSpeed — the foundation of our distributed training stack
• Accelerate — the launcher that ties everything together
• PEFT — the LoRA implementation behind lora_rank=128 and critic_lora
• xFormers — memory-efficient attention
• Flash-Attention — fast and memory-efficient exact attention (FA2 / FA3)
• Triton — custom kernels for RMSNorm, RoPE, and tiled linears (gigaworld/modules/gigaworld_kernels/)

🗃️ Data & Annotation

• Qwen3-VL — dense long-caption generation from cam_high videos
• Depth Anything V2 — monocular depth estimation for all three camera views
• LeRobot — the LeRobot-style raw data layout our preprocessing pipeline consumes
• Open X-Embodiment & AgiBot — large-scale robot demonstration datasets

🦿 Robotics, Visualization & Tooling

• Three.js — the WebGL renderer behind the URDF Viewer and camera-calibration tool
• Three.js + URDFLoader — URDF/STL loading and forward kinematics
• WandB — experiment tracking (offline-by-default in this repo)
• Pandas, NumPy, Pillow, OpenCV — the daily workhorses

🎬 Predecessors & Inspiration

• VideoCrafter, CogVideoX, Open-Sora, LTX-Video — for showing us what open-source video generation can look like
• Wan-Video / Fun-1.1-1.3B-InP — base model artifacts
• DMD / DMD2 — the diffusion distillation theory behind our Stage-2 acceleration

🌟 Community

A heartfelt thank you to everyone who has filed an issue, opened a PR, shared a workflow, or simply starred the project. Open source is a relay race — we are proud to hand the baton forward.

If you find GigaWorld-1 useful, please consider ⭐ starring the repo and citing the paper (see below).

---

📖 Citation

@article{team2026gigaworld,
  title={GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation},
  author={Team, GigaWorld and Ma, Angyuan and Wang, Boyuan and Li, Bohan and Ni, Chaojun and Li, Guo and Huang, Guan and Zhao, Guosheng and Li, Hao and Li, Hengtao and others},
  journal={arXiv preprint arXiv:2607.02642},
  year={2026}
}

---