| tags: | |
| - audio | |
| - synthesizer | |
| - sound-design | |
| license: apache-2.0 | |
| # Vital Decoder V1 + JEPA V7 | |
| Standalone baseline decoder for Vital JEPA V7. It predicts a category-balanced 64-preset structural codebook asset and sparse residual edits over 40 supported controls. Training uses similarity-weighted soft codebook targets so structurally close presets share credit. | |
| The selected decoder is step 1,000, chosen by held-out validation before asset-ID overfitting began. The bundled `codebook/` contains the complete preset asset bank, so inference does not require the original 7,522-preset database. | |
| Held-out test metrics and the exact config are included. This remains a baseline: the residual control vocabulary and one-shot-to-preset ambiguity are known fidelity limits. | |