Direct comparison
The panel stays fixed.
Every source clip, noise seed, normalization rule, and frozen 10-codebook codec matches the prior page. Enhancement uses Heun-4: two predictor/corrector intervals and exactly four velocity evaluations. “Lineage seen” and “prior unseen” preserve the earlier overfit-lineage panel labels; they are not train/test labels for this scratch run. This is not a strictly recording-disjoint DAC20k test: seven clips in each cohort share a recording with DAC20k train, and one lineage-seen clip is an exact DAC20k train path.
Ground truthNormalized source
NoisyCorrupt → frozen codec
EnhancedHeun-4 → frozen decoder
Loading the evaluation panel…