The simplest end-to-end check of an installation is the identity sanity baseline, which copies each source clip as its own prediction:

hf download ViTeX-Bench/ViTeX-Dataset --repo-type dataset --include "eval/*" --local-dir data
python benchmark/make_identity_baseline.py \
    --records  data/eval/parsed_records.json \
    --src_dir  data/eval/original_videos \
    --out_dir  baseline_output_videos/identity
bash scripts/run_benchmark.sh identity

Identity should give PSNR_loc = 100, SSIM_loc = 1 and LPIPS_loc = 0 on every clip, and its aggregates should match the Source video row of results/summary.tsv. If they do not, the pipeline is not installed correctly.

To produce predictions with the reference model, see vitex_edit/README.md.