reseed-ssim-references

Re-seeding SSIM references involves backing up current videos, regenerating new outputs on Modal, reviewing quality side-by-side, and uploading verified replacements to ensure accurate video quality similarity benchmarks during regression testing.

You need to re-seed video quality references when model code or hyperparameter modifications cause performance drift, rendering existing reference videos invalid for accurate similarity benchmarking and regression tracking.

Yes, targeted re-seeding focuses on a specific test file and model ID pairing to safely replace outdated reference videos on HF datasets with newly generated outputs for precise model validation.

The re-seeding process safely backs up existing reference videos, regenerates new ones on Modal, performs side-by-side quality comparison for user review, and executes scoped uploads of revised reference videos.

Yes, the workflow requires the hf dependency to perform scoped uploads of revised reference videos to HF datasets, ensuring updated similarity benchmarks are correctly stored for model validation.