InWorld: selected S1-S5 models
Action-conditioned latent video world model with Full-ST chunk-causal attention, Rich32 actions and a rolling diffusion-forcing/shortcut sampler.
Training code, launchers, engine collection code and tests: mingtiannihao/Inworld/training.
Released versions
| Stage | Selected local update | Files |
|---|---|---|
| S1 | 3000 | s1-u3000 |
| S2 | 10000 | s2-u10000 |
| S3 | 10000 | s3-u10000 |
| S4 | 13000 | s4-u13000 |
| S5 | 7000 | s5-u7000 |
S5 u7000 is the selected final checkpoint. Each stage directory contains
model.safetensors, its exact architecture config.json, and provenance.json.
Weights preserve the original mixed BF16/FP32 parameter precision. The release
contains five model-only files, approximately 26.3 GB total. Distributed optimizer
states and rank RNG files remain in the source workspace.
Load
Install the training package and dependencies following its README, then:
from huggingface_hub import hf_hub_download
from wm.release import load_release_model
repo = "yuduan/inworld"
prefix = "stages/s5-u7000"
config = hf_hub_download(repo, f"{prefix}/config.json")
weights = hf_hub_download(repo, f"{prefix}/model.safetensors")
model = load_release_model(config, weights, device="cpu")
This is a custom PyTorch model. It expects VAE latents, Rich32 actions, denoising progress and step-size conditions. RGB generation also requires the Wan2.2 VAE and the rolling sampler from the code repository. The VAE and training datasets are obtained separately.
Configuration and provenance
The backbone has 32 blocks, width 2048 and 32 heads. Inputs use 48-channel VAE latents and Rich32 actions. S5 long-clip sampling retains at most 13 completed chunks plus 12 active chunks and one persistent first-frame anchor; short replay retains at most 6 completed chunks. One chunk is 4 latent frames, or approximately 0.8 seconds at 20 FPS with VAE temporal stride 4. The selected route uses recent contiguous history; legacy sparse-memory branches are disabled.
release_manifest.json, provenance.json and SHA256SUMS identify the source
steps, tensor layouts and exact weight files. Original absolute source paths
inside safetensors metadata are provenance and are not required at load time.
Selected update numbers are stage-local, not one continuous global step counter.
Validation videos
The validation archive preserves 4,524 original local video paths as 2,795 unique video files (7.68 GB). It includes all local S1-S5 validation/development videos, comparisons and labeled reference previews, including videos retained only in local W&B media. CSV index, JSONL index, and manifest preserve every original run/source path and map duplicates to the uploaded object. Older experimental clips are identified separately from the selected S5 version.
Game engine code
The collection scripts, scene/action configuration, launchers and validators are included in the GitHub training directory. See engine setup and code map. The original engine training dataset is separate from this code/video release.
Selected S5 u7000 / K12 examples
These are archived development-validation clips; the index records their exact source paths.
s5_engine_controlled_navigation_00
s5_vpt_turn_left_00
slow193_vpt_move_button_00