Physical Evolution Supervision for Embodied World Models
Project Page · Quick Start · Reproduction · BibTeX
EVEWorld combines restoration supervision for target-instance consistency with temporal alignment for cross-frame consistency.
Visual plausibility does not guarantee physically consistent target evolution. In embodied video world models, the manipulated target may be spuriously duplicated or disappear during a rollout, while count-preserving deformation or identity drift can still occur across adjacent frames.
EVEWorld introduces two complementary forms of evolution supervision: Instance-Guided Restoration (IGR) restores clean demonstrations from count-perturbed inputs to promote target-instance consistency, while Temporal Instance Alignment (TIA) aligns target representations across adjacent frames to promote cross-frame consistency. We further introduce Model Laziness Rate (MLR), a rollout-level, occlusion-aware and persistence-aware diagnostic for persistent target-instance violations.
Quantitative and qualitative results are available on the project page and in the paper.
Our FlowWAM-based EVEWorld submission, Supervision_WM, achieves an EWMScore-P of 70.17 on WorldArena 2.0 Track 1 — Simulator Video Quality, ranking 6th in JEPA Similarity and 17th overall.
70.17 EWMScore-P · 6th in JEPA Similarity · 17th overall
View the official WorldArena 2.0 leaderboard ↗
- IGR training pipeline
- TIA alignment module
- MLR evaluation protocol
- DreamGen / WorldArena / EWMBench / RoboTwin evaluation entry points
- paper-protocol reproduction entry points
- the item-level evaluation-row schema in
docs/item_level.md - environment and third-party-backbone setup scripts
- the project page and its qualitative media (served from the
gh-pagesbranch)
conda env create -f envs/gigaworld.yaml
conda activate gigaworld
pip install -e ".[train,eval]"
python scripts/setup/check_environment.pypython scripts/setup/check_environment.py exits non-zero only when a core item is missing; optional items are reported and otherwise ignored unless --strict is given.
conda env create -f envs/evaluation.yaml
conda activate eveworld-eval
pip install -e ".[eval]"conda env create -f envs/flowwam.yaml
conda activate flowwam
pip install -e ".[train,eval]"The first end-to-end reproduction is bash scripts/reproduce/table1_dreamgen.sh; the full artifact map is in Reproducing the paper.
The main paper setting uses 8× NVIDIA H20Z GPUs with an effective batch size of 64. Reproduction requirements differ across evaluation-only, GigaWorld, and FlowWAM workflows; see docs/installation.md and docs/reproduction.md for per-task requirements.
EVEWorld integrates with GigaWorld-0 for the main DreamGen setting and FlowWAM for the cross-backbone RoboTwin experiments. Upstream source trees and weights are not redistributed in this repository.
bash scripts/setup/clone_gigaworld.sh # -> third_party/giga-world-0
bash scripts/setup/clone_flowwam.sh # -> third_party/FlowWAMGigaWorld-0 comes from https://github.com/open-gigaai/giga-world-0. Both scripts accept a destination argument and a --ref override, and both checkout directories are gitignored. clone_gigaworld.sh also installs the giga-train and giga-datasets packages, which are not on PyPI. FlowWAM training starts from the released Stage-1 checkpoint on top of the Wan2.2-TI2V-5B base; both are fetched by bash scripts/setup/download_models.sh. See docs/third_party.md and third_party/README.md for the weight table and licenses.
Instance-Guided Restoration (IGR). IGR constructs count-edited demonstrations by inserting an additional target instance and trains the model to restore the clean latent representation. Restoration-related regions receive 3× reconstruction weight, with the spatial weight map normalized to unit mean.
Temporal Instance Alignment (TIA). TIA operates at a selected transformer layer, performs local cross-frame target matching, and transports matched previous-frame features into the current representation through a bounded residual. A correspondence objective favors the true next-frame target location.
The two objectives are optimized jointly. At inference, EVEWorld requires only the initial image and instruction; external target localization is not required.
See docs/training.md for training schedules and implementation details.
Each paper artifact has a script under scripts/reproduce/:
| Paper artifact | Command | Output |
|---|---|---|
| Table 1 · DreamGenBench | bash scripts/reproduce/table1_dreamgen.sh |
outputs/evaluation/item_level/dreamgen/ |
| Table 3 · WorldArena 1.0 | bash scripts/reproduce/table3_worldarena.sh |
outputs/evaluation/item_level/worldarena/ |
| Table 4 · EWMBench (AgiBot) | bash scripts/reproduce/table4_ewmbench.sh |
outputs/evaluation/item_level/ewmbench/ |
| Table 5 · RoboTwin (FlowWAM) | bash scripts/reproduce/table5_robotwin.sh |
outputs/evaluation/item_level/robotwin/table5.json |
| Table 6 · Component ablation | bash scripts/reproduce/table6_ablation.sh |
outputs/evaluation/item_level/dreamgen/ |
| Figure 4 · CFG grid | bash scripts/reproduce/figure4_cfg.sh |
grid cells under ${GRID_ROOT}; summaries under outputs/evaluation/item_level/dreamgen/ |
See docs/reproduction.md for inputs, compute requirements, and per-artifact protocols. Quantitative and qualitative results are on the project page and in the paper; the protocols, metrics, and judge details are in docs/evaluation.md.
EVEWorld/
├── assets/ README figures; the project-page media lives in the gh-pages branch
├── configs/ training, evaluation, and ablation run configurations
├── data/ splits, per-clip metadata, and judge prompts (datasets stay outside the repo)
├── docs/ installation, data, training, inference, evaluation, reproduction, checkpoints, third party
├── envs/ conda environment files (gigaworld, flowwam, eveworld-eval)
├── experiments/ alternative supervision designs and the analyses behind the paper figures
├── outputs/ evaluation outputs of the reproduction scripts (gitignored)
├── scripts/ setup, prepare, train, inference, evaluate, and reproduce entry points
├── src/eveworld/ the library: IGR, TIA, data tooling, integrations, and evaluation
├── tests/ unit tests for the IGR, TIA, and MLR components
└── third_party/ clone instructions for the two backbones (checkouts are gitignored)
| Document | Covers |
|---|---|
docs/installation.md |
environments, extras, hardware, and environment variables |
docs/data_preparation.md |
datasets, splits, and the scripts/prepare/ converters |
docs/training.md |
training recipes, schedules, and commands |
docs/inference.md |
inference entry points and generation settings |
docs/evaluation.md |
MLR, judges, and the benchmark harnesses |
docs/item_level.md |
the per-item evaluation-row schema |
docs/reproduction.md |
inputs, compute, and per-artifact protocols |
docs/checkpoints.md |
downloader flags, landing paths, and checkpoint layout |
docs/third_party.md |
backbones, weights, and licenses |
The author list remains anonymous; the release links the project page and this repository.
@software{eveworld2026,
title = "{EVEWorld}: Physical Evolution Supervision for Embodied World Models",
author = "{Anonymous Author(s)}",
year = {2026},
url = "{https://behappya.github.io/EVEWorld/}",
note = "Code and project page released for anonymous review."
}The machine-readable entry is CITATION.cff.
The code is released under the Apache-2.0 License. Third-party components — GigaWorld-0 / GigaModels, FlowWAM, GroundingDINO, SAM2, RAFT, LPIPS, WorldArena, DreamGenBench, EWMBench, and PBench — keep their own licenses; see docs/third_party.md and third_party/README.md.
EVEWorld builds on the GigaWorld-0 and FlowWAM backbones, and relies on GroundingDINO and SAM2 for target localization and tracking. See docs/third_party.md for details and licenses.
Contributions are welcome; see CONTRIBUTING.md. The unit tests run with pytest -q, and the formatting and lint hooks with pre-commit run --all-files; the tests workflow runs both on every push.