license: apache-2.0
tags:
- world-model
- minecraft
- video-generation
- causal-forcing
- action-conditioned
ForgeWM Checkpoints
Model checkpoints for ForgeWM, a reproducible progressive-causal training recipe for few-step, action-conditioned video world models.
- 🌐 Project page: https://asdfo123.github.io/ForgeWM
- 💻 Code: https://github.com/asdfo123/ForgeWM
- 📦 Training data:
ForgeWM/ForgeWM-data
Checkpoints
The four Minecraft stages form a single, self-consistent lineage: stage3 is
distilled from stage2, which is distilled from stage1. stage0 is the
bidirectional teacher the DMD stages score against.
| File | Stage | Description |
|---|---|---|
stage0/model.pt |
Stage 0 | Bidirectional SFT for domain adaptation off the MG2 base (4,000 updates). Also the frozen real-denoiser teacher_ckpt that Stage 3 scores against. |
stage1/model.pt |
Stage 1 | Teacher-forcing causal AR, trained from the MG2 base for 20,000 updates at lr 2e-5. Stage 0 and Stage 1 are siblings off the base model. |
stage2/model.pt |
Stage 2 | Consistency distillation of that Stage 1 (6,000 updates). |
stage3/model.pt |
Stage 3 | ForgeWM-4 — the 4-step DMD student distilled from stage2. This is the default real-time model. |
1step/model.pt |
Few-step | ForgeWM-1 — 1-step DMD student. Shares Stage 0–2; uses the first-chunk FFE schedule. |
2step/model.pt |
Few-step | ForgeWM-2 — 2-step DMD student. Shares Stage 0–2; uses the first-chunk FFE schedule. |
crossfps/model.pt |
CrossFPS | Cross-domain inference checkpoint: the same recipe ported to a gamepad-driven FPS (4-D dual-stick, 6 buttons). A separate lineage, not a multi-domain model. |
Running Stage 3 (ForgeWM-4)
stage3 was trained with an unconditional train-time KV-cache refresh, and it
must be run with the same sliding-window attention it was trained under —
forward the top-level local_attn_size: 6 / sink_size: 0 into the inference
wrapper (the shipped configs/stage3_dmd.yaml does this). Running it with full
causal attention silently evaluates it off-distribution.
The 1-/2-step students (1step, 2step) run their budget-matched schedule
everywhere and the full 4-step schedule on chunk 0 (first-chunk FFE); the
matching configs/stage3_dmd_1step.yaml / stage3_dmd_2step.yaml set this up.
Usage
# Download all checkpoints
huggingface-cli download ForgeWM/ForgeWM --local-dir ./ckpts --repo-type model
# Inference (default 4-step recipe, ForgeWM-4)
python inference.py \
--config_path configs/stage3_dmd.yaml \
--checkpoint_path ckpts/stage3/model.pt \
--image_path demo_images/forest.png \
--action_type forward
Citation
@misc{li2026forgewm,
title = {ForgeWM: Progressive Causal Training for Few-Step
Action-Conditioned Video World Models},
author = {Xinye Li and Lingshuai Lin and Lei Wang and Liuzhou Zhang and
Jialin Cui and Qingshan Li and Guanchu Wang and Qingbin Liu and
Xi Chen and Jiang Bian and Wai Lam},
year = {2026},
url = {https://github.com/asdfo123/ForgeWM}
}
Acknowledgements
- Base model: Matrix-Game 2
- Minecraft training data: GameFactory
- CrossFPS training data: SCOPE
- Distillation recipe: Causal Forcing, building on Self-Forcing and CausVid