You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

JAIS physical graph experiments

Wan mask training comparison

Scene 008669. Watch the lavender sphere move left after the edit. Both Wan models receive the same predicted masks and sampling noise. From left to right, the panels show the factual video, the true counterfactual, Wan trained on ground-truth masks, and Wan trained on attention-predicted masks.

Play the predicted-mask comparison

With ground-truth masks supplied to both models, the first panel shows the control mask. The other three panels keep the same order.

Play the ground-truth-mask comparison · Browse all 15 comparison scenes

Counterfactual video comparison on the expanded JAIS dataset

Two Wan models were finetuned separately, one with ground-truth masks and one with masks from the retrained attention propagator. Both generated videos from the same fixed inputs and sampling noise on 16 held-out scenes. The gallery includes all 16 examples, the oracle-mask comparison, and post-edit metrics for moving objects and background. Full comparison and videos

The first registered scene is 011186. From left to right the panels show the factual video, the true counterfactual, Wan trained on ground-truth masks, and Wan trained on predicted masks.

Counterfactual video comparison for scene 011186

Play the matched comparison for scene 011186

Matched video results

With the same attention-predicted masks, GT-trained Wan reached 19.03 dB post-edit foreground PSNR and attention-trained Wan reached 18.47 dB across five scenes from each data version. Across both panels the experiment generated 92 clips from 23 unique test scenes. Results and visual observations · Earlier Wan comparison

Wan videos across three generation batches

The batch comparison gives v1, v2 and v3 five held-out scenes each. Both Wan checkpoints use identical inputs and noise. The full gallery includes a ground-truth control comparison for every scene and separate scores for each batch. 15-scene gallery and per-batch results

Video comparison for v1 scene 000772

Play v1 scene 000772

Downloads last month
138