language:
- en
license: other
license_name: ltx-2-community
license_link: LICENSE
pipeline_tag: image-to-video
CausalWMv1
The current released checkpoint is tailored for PAI-Bench-G. Additional checkpoints will be released soon.
CausalWMv1 takes one RGB observation and a text instruction, then generates optical flow, camera-frame XYZ pointmaps, and future RGB video through a stage-causal chain.
- Code: AetherLabsAI/CausalWM
- Paper: CausalWM: Causal Chain-of-Thought Reasoning for Embodied World Model
- Checkpoint:
CausalWMv1.safetensors
Files and requirements
The checkpoint contains the complete CoT Transformer state in BF16. It is not a
standalone pipeline: inference additionally needs ltx-2.3-22b-dev.safetensors
from Lightricks/LTX-2.3 and the
Gemma-3-12B text encoder.
See model-manifest.json for the file size, SHA-256, tensor count, and runtime
metadata. Verify the download using sha256sum -c SHA256SUMS.
Inference
Follow the repository installation instructions for TI2V inference, then run:
python inference.py \
--image /path/to/first_frame.png \
--prompt "The robot arm picks up the red block and places it in the box." \
--checkpoint /path/to/CausalWMv1.safetensors \
--base-ckpt /path/to/ltx-2.3-22b-dev.safetensors \
--text-encoder-dir /path/to/gemma-3-12b \
--out-dir outputs/demo
Default settings are 121 frames, 640×480, 16 FPS, four steps per stage, guidance 1.0, and seed 42. The input frame is resized without cropping. Pointmaps use a relative camera-frame scale normalized by the first frame's median depth, not metres. Generated motion and geometry are predictions rather than measured ground truth.
License and attribution
See LICENSE and NOTICE for the LTX-2 license and attribution carried with this
release. Base model and text encoder downloads remain subject to their respective
terms. This package contains inference weights, not optimizer or training state.
Citation
If CausalWM is useful for your research, please cite our paper:
@misc{xu2026causalwmcausalchainofthoughtreasoning,
title={CausalWM: Causal Chain-of-Thought Reasoning for Embodied World Model},
author={Ziming Xu and Shuang Liang and Ruobing Han and Ziqiao Xi and Mingxing Rao and Kun Zhou and Zijun Zhang and Yuchen Yan and Yufan Wei and Junbo Huang and Yifei Shao and Fang Nan and Biwei Huang},
year={2026},
eprint={2609.23184},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2609.23184},
}