| Training | Original | Reconstructed |
|---|---|---|
| First trainings | ▶ Original | ▶ Reconstructed |
| 2025-11-29/03-27-25 | ▶ Original | ▶ Reconstructed |
At this point, the generation of the tokens for the RVQ is done autoregressively one at a time, without any hierarchical autoregressive modeling, the tokens are then flattened to compute a MSE Loss. the next iteration will implement RQ-Transformer (temporal+depth transformers) for hierarchical autoregressive modeling
Pull Encodec repo:
git submodule update --init