RAVE architecture configs trade reconstruction quality, GPU memory, and control modality

RAVE ships multiple gin-based architecture configurations. v2 is the default improved continuous model (16 GB GPU min); v2_small is faster and suited to stationary timbres (8 GB); v3 adds Snake activation and Adaptive Instance Normalization for style transfer (32 GB); discrete mimics SoundStream/EnCodec with a codebook latent (18 GB); raspberry targets embedded realtime inference (5 GB). The continuous configs (v1/v2/v3) produce a smooth latent space suitable for latent manipulation; the discrete config uses quantized indices — better reconstruction but the latent is not continuous, so slider-based control is unavailable. Configs are composable: --config discrete --config causal enables causal discrete mode.

Examples

rave train --config v2_small --db_path /data --out_path /out --name drums for a compact drum timbre model. rave train --config v3 for style transfer capability.

Assessment

A performer wants to control 8 continuous latent dimensions live via MIDI CC. Which RAVE config family should they avoid and why?