Looking-Ahead WISA latent demos

Pre-encoded example data for Looking-Ahead TF and FCP recipes: 100 distinct WISA video identities, each encoded for two Wan families, giving 200 latent records. These are loader-format examples, not the full historical training corpus or an evaluation benchmark. Source videos are not included.

Contents

Family	Compatible recipes	Source A	Source B	Disk [C,T,H,W]	Loader [T,C,H,W]
wan21_t2v_1.3b	TF/FCP 1.3B	50	50	[16,22,54,96]	[22,16,54,96]
wan22_ti2v_5b	TF/FCP 5B	50	50	[48,31,36,64]	[31,48,36,64]

A/B membership is disjoint. Both families have identical video identities, source membership and captions; TF/FCP share their family's inputs. Use the matching family: these two VAE encodings cannot be exchanged by reshaping.

wan21_t2v_1.3b/             # same structure under wan22_ti2v_5b/
  source_a/
    annotations.jsonl
    latent_shards/latent_<shard>.tar
  source_b/
    annotations.jsonl
    latent_shards/latent_<shard>.tar

Each UTF-8 JSONL record supplies shard, clip_name and global_caption_en. The tar path is latent_shards/latent_ + shard + .tar; a tensor member is <clip_name>.latent.pth. Captions come from JSONL. The production loader converts [C,T,H,W] into float32 [T,C,H,W]. Use the project's latent WebDataset loader; this package is not a table for automatic datasets.load_dataset loading.

Download

Download the dataset demo from the companion code checkout root:

HF_DATASET_REPO=TencentARC/Looking-Ahead
hf download "$HF_DATASET_REPO" --repo-type dataset \
  --local-dir release_assets/dataset_demo

To fetch one family, add --include 'wan21_t2v_1.3b/*' (or the 5B folder). Keep both source_a/ and source_b/: public configs already name these paths under release_assets/dataset_demo/. The complete dataset is about 2.0 GiB. Trained project weights and upstream Wan components are acquired separately.

Load a source

With the code checkout's ordinary uv environment installed, run from its root:

uv run python - <<'PYCODE'
from src.dataset.tensor import PytorchTensorWebDataset

dataset = PytorchTensorWebDataset(
    data_path="release_assets/dataset_demo/wan21_t2v_1.3b/source_a/latent_shards",
    annotation_path="release_assets/dataset_demo/wan21_t2v_1.3b/source_a/annotations.jsonl",
    shuffle=0, shuffle_initial=0, shardshuffle=False, resampled=False, seed=42,
)
sample = next(iter(dataset))
print(sample["clean_latent"].shape)  # torch.Size([22, 16, 54, 96])
print(sample["prompts"])
PYCODE

For 5B, replace the family path; the returned shape is [31,48,36,64].

Optional validation and training

Run the manual validator from the code checkout root:

uv run python tools/validate_latent_webdataset.py \
  --data-path release_assets/dataset_demo/wan21_t2v_1.3b/source_a/latent_shards \
  --annotation-path release_assets/dataset_demo/wan21_t2v_1.3b/source_a/annotations.jsonl \
  --config tf_wan21_t2v_1.3b --all-shards --all-samples

Repeat for B and the other family as needed. Ordinary training performs no automatic complete-dataset scan. Once upstream Wan components and the code environment are ready, start a matching recipe:

bash run.sh train tf_wan21_t2v_1.3b --disable_wandb

Other names are fcp_wan21_t2v_1.3b, tf_wan22_ti2v_5b and fcp_wan22_ti2v_5b. Training on this tiny demo does not reproduce historical training or paper results.

Provenance, validation and license

Source: qihoo360/WISA-80K. The lexicographically first 100 common annotated video IDs were selected: first 50 for A, next 50 for B. Existing Wan-family tensors and captions were retained while selected records were repacked.

All 200 records were loaded with the actual production loader during this preparation: shapes, float32 output, finite values, captions, unique IDs, disjoint A/B membership and cross-family agreement passed. See VALIDATION.json, manifest.json and each family's provenance.json. No all-record VAE decode or quality evaluation is claimed. After downloading all public files, verify them with sha256sum -c SHA256SUMS.

The official upstream card declares Apache-2.0. See LICENSE, NOTICE and WISA_UPSTREAM_README.md for attribution. The pinned upstream revision identifies the license/attribution card inspected for delivery; the original download revision of the pre-existing video encodings was not recorded. This package retains the upstream declaration.