FAR checkpoints

Checkpoint bundles for FAR, a latent-diffusion world model with a learned, action-conditioned retrieval memory, and its baselines, on the LoopNav, SoundSpaces and AI2-THOR corpora.

• Paper: Learning What to Recall: Adaptive Multi-Cue Episodic Memory for World Models
• Code: https://github.com/sony/far
• Project page: https://1202kbs.github.io/FAR-Project-Page/

Layout

The repo mirrors the code's models/ directory, so a download lands exactly where the configs, launchers and notebooks look:

<corpus>/<arm>/config.yaml                  # rebuilds the model (Hydra)
<corpus>/<arm>/checkpoints/<step>.pth.tar   # EMA generator weights + memory state + step (inference only)
oasis_500m_vit_vae.pth                      # ViT-VAE tokenizer for LoopNav latents
agent_cell_detector.pt                      # latent agent-cell detector for the AI2-THOR-dyn probe
bundles.json                                # this listing with sizes and SHA-256

Bundles hold what inference needs (EMA generator, the trained or frozen memory state, the training step) and cannot resume training. The SDXL VAE used for the SoundSpaces and AI2-THOR corpora is fetched from madebyollin/sdxl-vae-fp16-fix automatically.

Download

python scripts/download_release.py                       # everything, ~13 GB
python scripts/download_release.py --corpus ai2thor_v3   # one corpus
python scripts/download_release.py --arm loopnav/far_multicue

or huggingface-cli download 1202kbs/FAR-Checkpoints --local-dir models.

Bundles

"Temporal" and "WorldMem" are the recency and field-of-view retrieval baselines, "LongLive-RAG" the content-query retrieval baseline; the FAR arms differ in the cues the retriever fuses (metadata, visual, multi-cue); the two LoopNav ablations complete the paper's LoopNav ablation table. The retriever bundles are the contrastively pre-trained encoders the FAR launchers start from; evaluation does not need them (each FAR bundle already carries its trained retriever).

License

CC BY-NC 4.0. The generator and diffusion code these weights belong to are adapted from Navigation World Models and DiT (Meta Platforms, CC BY-NC 4.0); see the code repository's THIRD_PARTY_NOTICES.md.

The SoundSpaces bundles (soundspaces_v1/*, soundspaces_v2/*) were trained on episodes rendered from Matterport3D scenes, so their use is additionally subject to the Matterport3D Terms of Use (non-commercial academic research).

Citation

@article{kim2026far,
  title   = {Learning What to Recall: Adaptive Multi-Cue Episodic Memory for World Models},
  author  = {Kim, Beomsu and Lai, Chieh-Hsin and Nguyen, Bac and Bar, Amir and Ye, Jong Chul and Mitsufuji, Yuki},
  journal = {arXiv preprint arXiv:2609.34677},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.34677}
}