open-gigaai/giga-world-1

TL;DR: GigaWorld-1 is an open stack (Apache-2.0) for training and running video world models whose stated purpose is evaluating robot policies without real-robot rollouts: feed in a first frame plus a control signal, get a 33-second, three-camera-view video prediction of what the robot would do. It's less a new idea than an unusually complete release of the plumbing — Stage-1 weights (1.3B "Nano" and 5B "Pro"), training recipes, data preprocessing, LoRA merge tools, and inference scripts — for a problem (cheap policy eval) that's a genuine bottleneck. The catch: the README contains zero quantitative evidence that world-model scores track real policy performance, and the benchmark (WMBench), distilled weights, and RL post-training are all only partially released or "coming soon."

What it is and why it matters

Policy evaluation is the quiet bottleneck of robot learning: real rollouts are slow and expensive, and simulators diverge from reality. The pitch here is to use a controllable video-generation world model as the evaluator — condition on the current scene and the policy's intended trajectory, roll the video forward, and score the result with fine-grained metrics (WMBench claims 15 metrics plus VLM judging, backed by a CVPR 2026 workshop challenge and public leaderboard).

The models themselves are fine-tunes of Alibaba's Wan video backbones: Nano is Wan 2.1 1.3B FunControl, Pro is Wan 2.2 5B FunControl. Output is 480×1920 — three 640-wide camera views tiled side by side — at 10 FPS, up to 330 frames (~33 s). Conditioning comes from a text prompt, an optional first frame (i2v vs t2v), and an optional control video encoding camera geometry as Plücker/ray maps.

How the pipeline works

Training is two stages. Stage-1 is controllable post-training on robot data (LeRobot-format episodes converted to their format, with Qwen3-VL dense captions and Depth Anything V2 depth for all views, latents precomputed offline). Stage-2 is DMD2 distillation, compressing the denoising loop from 20 steps to 4–6 using a frozen real-score model and a critic_lora — the standard recipe for making diffusion rollouts fast enough to be a practical evaluator.

A notably practical claim: adapting to a new robot domain needs no code changes — just format your data and run Stage-1 for a few thousand steps, typically finishing within a day on 8 GPUs.

What's actually there — and what isn't

Released: Stage-1 Nano and Pro weights (HF + ModelScope), Stage-1 and Stage-2 training code with DeepSpeed ZeRO-2/3 configs, i2v/t2v inference scripts, a LeRobot-style preprocessing pipeline with toy data, a web-based URDF/calibration viewer, and LoRA-merge/checkpoint tools. There's a technical report PDF and a CVPR 2026 challenge dataset and leaderboard.

Not released: Stage-2 distilled checkpoints (so the fast 4–6-step inference exists only as training code for now), 3D RL post-training, other-domain weights, the inference acceleration framework, and most of WMBench itself.

The bigger gap is evidence. The README gives no numbers — no FVD, no controllability metrics, and crucially no correlation between world-model evaluation scores and real-robot policy success. That's the entire premise of the project, and you'll need the technical report (arXiv 2607.02642, PDF-only at the moment) or the leaderboard to judge it. Treat "roadmap" in the title as accurate: this is infrastructure plus a promise, not a validated evaluator yet.

Practicalities

Production setup is a single 8×H20 or 8×A100 node; inference of Nano (1.3B) should be feasible on consumer GPUs with VAE tiling and memory-saving settings. Linux only, CUDA 12.x, PyTorch 2.10, a custom-modified diffusers fork in thirdparty/, optional FlashAttention 2/3, and custom Triton kernels for RMSNorm/RoPE. Apache-2.0 — but note the underlying Wan backbones carry their own license terms.

Try it

pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 \
  --index-url https://download.pytorch.org/whl/cu126
bash install.sh

# weights + toy data
bash tools/download_tool/download_giga_world.sh --platform hf --target model   --output-dir ./downloads
bash tools/download_tool/download_giga_world.sh --platform hf --target toydata --output-dir ./downloads
mkdir -p model example
cp -r ./downloads/Giga-World-1/* ./model/
cp -r ./downloads/Giga-World-1-Toydata/* ./example/

# Nano i2v rollout (first frame + control video + prompt)
bash scripts/infer/run_infer_nano_i2v.sh

Weights: huggingface.co/open-gigaai/Giga-World-1. Watch the repo's releases — the distilled checkpoints and full WMBench are where this becomes (or fails to become) a practical policy evaluator.