Embedders and pretrains

Inference embedders

Open Embedder Settings below Audio Processing. Auto reads stored model metadata and uses ContentVec when no embedder is recorded. If the publisher specifies an encoder, use that encoder. A retrieval index needs compatible features from the same experiment. Leave retrieval disabled when the pairing is uncertain.

Available inference encoders include ContentVec, Spin, Spin-v2 and Chinese, Japanese and Korean HuBERT Base. Open Other → Download embedders to install optional models one at a time. For custom HuBERT assets, select custom, name the folder, provide the matching weights and configuration files (.bin and .json), and install them together.

Training compatibility

Training exposes ContentVec and Spin-v2. Changing the encoder requires fresh features and a separate experiment name. Match your generator and discriminator pair to the vocoder, training sample rate, RVC generation and pitch configuration, and follow the pair publisher's encoder requirements. Ilaria RVC 2.1.0 supports RefineGAN training specifically at 32 kHz with RVC V2 and pitch guidance. The downloader may list other rates for use elsewhere.

Get a G/D pair

Under Other → Download pretrained models, use a repository or pretrained archive, or enter a link to the G checkpoint and its matched D checkpoint if discovery needs help. The Recommended pretraineds picker shows available catalog variants and their sample rates. Check the chosen variant before installing both files. Return to Training, refresh the list, and select the installed pair. If using random initialization intentionally, choose None. Keep inference models and training pretrains separate.

If training exits immediately after selecting a pair, verify both files and the saved experiment configuration before changing batch size or starting a new run.