CamxTime_new

Precomputed video latents, camera parameters, motion frame indices, timestamps, captions, and a shared text embedding bank. Tensor files are uploaded byte-for-byte.

Split	Safetensors files
train	25,500
eval	256
text_bank	2,311

Train covers 417 scenes and four trajectories per scene. The 1,569 moving-camera trajectories each contain source, backward, zigzag, slowmo, and 12 bullettime clips. The 99 static-camera trajectories contain only the four non-bullettime patterns. Bullettime freezes motion at original frame 0, 10, ..., 110 while camera indices advance. All 18,828 training bullettime files have 81 frames at 832x480 and BF16 Wan2.1 VAE latents of shape [16, 21, 60, 104]. Frozen motion indices and timestamps were checked against the filename in all bullettime training files.

Each clip includes vae_latents, cam_c2w, cam_wm_c2w, intrinsic, intrinsic_wm, timestamps, source_cam_idx, source_motion_frame, and string metadata. The metadata text_key refers to a file in text_bank. Embedded original filesystem paths are provenance strings and are not needed to download the tensors. Evaluation files come from the existing CamxTime_latents evaluation split; their pattern inventory is independent of the training split described above.

Layout and download

Training files are grouped as train/Scene001/Scene001_traj01_source.safetensors to stay within Hub directory entry limits. eval/ and text_bank/ retain their flat layout. file-map.jsonl maps every original relative path to its repository path.

from huggingface_hub import snapshot_download
root = snapshot_download("HugMinjun/Camxtime", repo_type="dataset")

Use a recursive glob for train/*.safetensors, or use file-map.jsonl to restore the original flat train directory. Load tensor files with safetensors.safe_open; text embedding files are required in addition to the per-clip latent files.

No new license grant is asserted by this upload. Consult the original data's applicable terms before redistribution or downstream use.