Experimenting with video-based learning.
Videos seem less likely to lead to overfitting.
The voice hasn't been trained, so please bear with the mechanical sound and odd intonation.
gc_vg_minimaxh3_3397693_epoch_10.safetensors