kinema4d-smoke10

Ten encoded DROID clips used as the production-gate smoke for the Kinema4D trainer. This is not a research dataset — it is the minimum-viable real input that lets a single vendors/Kinema4D/finetune.py --train_steps 10 run end-to-end on real Wan2.1-I2V-14B + 4DNeX-LoRA weights, exercising the same code path that drives the full DROID training run.

The companion repository is samiazirar/contact_flow @ cluster/docker-reproducibility, where tests/run.sh real-data fetches this dataset and runs the smoke.

Layout

video_latents/<clip>.safetensors         # VAE-encoded RGB latents + CLIP image embedding
mask_video_latents/<clip>.safetensors    # VAE-encoded masked-RGB latents (I2V seed)
pointmap_latents/<clip>.pt               # VAE-encoded pointmap latents (contact-flow target)
mask_pointmap_latents/<clip>.pt          # VAE-encoded masked-pointmap latents
arm_masks_npy/<clip>.npy                 # rasterized contact-flow occupancy mask, uint8
images/<clip>.jpg                        # first-frame RGB (480x720) for I2V conditioning
cache_empty/                             # null-prompt CLIP embedding the trainer reuses
train_shuffled.txt                       # video_latents/<clip>.safetensors per line
train_img_shuffled.txt                   # images/<clip>.jpg per line

The two txt files point at real on-disk paths so the dataset class's path.is_file() check passes; the trainer strips .stem to recover the clip name.

Ten clips chosen to span four distinct DROID recordings × two camera extrinsics so the smoke catches per-recording and per-camera regressions without pulling the full ~300 GB shard.

Reproducibility

These latents were produced by the DROID encode-stage that ships with the Kinema4D vendor (vendors/Kinema4D/data_proc_full/encode_latents.py). The exact encoder commit is pinned in the trainer Dockerfile and the encoder reads RGB+pointmap mp4s plus per-frame contact-flow occupancy. Re-running the encoder against the same DROID clips reproduces these files bit-for-bit modulo VAE sampling noise (the encoder uses a fixed seed when called from the slurm wrappers).

Use

# inside the contact_flow checkout, branch cluster/docker-reproducibility
SIF_PATH=<cluster>/kinema4d.sif tests/run.sh real-data

tests/fetch_real_data.sh resolves this dataset via huggingface_hub, caches it under $SMOKE_DATA_ROOT/data_real/, and the slurm wrapper tests/smoke_real.slurm then runs vendors/Kinema4D/finetune.py --train_steps 10 against the cached copy.

License

The encoded latents are derived from the DROID dataset. Use of this subset is bound by the upstream DROID terms; this repository redistributes the encoded representation only for the purpose of running a deterministic smoke test against the Kinema4D trainer.