README.md

metadata

license: apache-2.0
task_categories:
  - text-to-video
pretty_name: One-Forcing Data
size_categories:
  - 1K<n<10K

One-Forcing Data

Compact clean-latent LMDB shards for One-Forcing training.

The original ODE trajectory rows had shape [num_denoising_steps, num_frames, channels, height, width]. This release keeps only the clean latent used by training, so each row is stored as [num_frames, channels, height, width].

Layout

ODE6KCausal_chunkwise_0/data.mdb
ODE6KCausal_chunkwise_0/lock.mdb
...
ODE6KCausal_chunkwise_14/data.mdb
ODE6KCausal_chunkwise_14/lock.mdb

Each shard uses the same LMDB key layout as the training code:

• latents_shape: stored shape for the shard, e.g. 434 21 16 60 104
• latents_source_shape: original full trajectory shape, e.g. 434 6 21 16 60 104
• latents_clean_only: 1
• latents_{i}_data: float16 clean latent row
• prompts_{i}_data: UTF-8 prompt string
• prompts_shape: number of prompts in the shard

Usage

Point One-Forcing training at the downloaded dataset directory:

torchrun --nproc_per_node=8 train.py \
  --config_path chunkwise_config.yaml \
  --data_path /path/to/One-Forcing-Data \
  --disable-wandb \
  --no_visualize

The expected data_prefix is ODE6KCausal_chunkwise.