kaimingyang/DVD_for_Dream.exe

Dream.exe DVD depth checkpoints

📄 Paper · 🤗 HF Daily Paper · 💻 Code · 🧪 Benchmark

These are the DVD depth checkpoints used by Dream.exe. We fine-tuned the official DVD model with LoRA on data from our Dream.exe benchmark, built using RoboCasa data and environments. The checkpoints support the video-to-trajectory stage of Dream.exe.

The shared model and UID-specific models are provided with their runtime configuration. Use the checkpoint for the input case UID; keep the UID names unchanged. The release contains one shared checkpoint and 116 UID-specific checkpoints: exactly 234 checkpoint/config files. Historical metadata-only case directories and training metadata are excluded.

Download and run

From the Dream.exe code repository root, download only the bundled quickstart case:

hf download kaimingyang/DVD_for_Dream.exe \
  --include 'DVD/lora/specific/rc_cheesybread_ep000001/*' \
  --local-dir checkpoints

For the full checkpoint family:

hf download kaimingyang/DVD_for_Dream.exe \
  --include 'DVD/lora/shared/*' 'DVD/lora/specific/*' \
  --local-dir checkpoints

The repository preserves the following paths:

DVD/lora/
├── shared/
│   ├── model.safetensors
│   └── model_config.yaml
└── specific/<uid>/
    ├── model.safetensors
    └── model_config.yaml

This downloads into checkpoints/DVD/lora/, which matches Dream.exe's default configuration. For another location, update roots.checkpoints and bindings.checkpoints.dvd_assets.path in your workspace configuration. The official DVD base and Wan runtime assets are installed separately by the commands in INSTALL.md. That guide contains the complete quickstart; the benchmark guide contains the complete dataset and reproduction commands.

These files are runtime fine-tuned checkpoints; do not assume they are portable adapter-only files loadable by an arbitrary LoRA library. Use the Dream.exe DVD backend and accompanying configuration. Training states and intermediate training checkpoints are outside this release.

Acknowledgements and terms

We thank the DVD authors for the official base model and the RoboCasa authors for the data and simulation environments used to construct our benchmark and fine-tuning data. See the DVD model and RoboCasa project for their original work. The weights follow DVD's CC BY-NC 4.0 non-commercial terms. Third-party data, code, and runtime assets retain their respective licenses.

Please cite Dream.exe and acknowledge DVD and RoboCasa when using these checkpoints. Dream.exe's citation is maintained in CITATION.cff.

Reproducibility

Pin the HF model revision for experiments. Dream.exe's packaged model manifest records the expected file hashes; verification must pass before execution. The directory names shared and specific/<uid> identify different model choices and should not be interchanged when reproducing a reported result.