ACE-Step 1.5

ACE-Step produces a full song — composition, vocals, instrumentation, and mix — from a text prompt or lyric sheet, the same kind of input as AutorunX's other song engines.

The difference is where it runs: instead of calling a paid third-party cloud API, it executes on AutorunX's own GPU pod, which is what keeps its per-song cost the lowest on the platform.

ACE-Step combines diffusion-based audio generation with a deep-compression autoencoder (derived from Sana's DCAE) and a lightweight linear transformer, letting it synthesize several minutes of coherent music far faster than LLM-token-based music models — publicly benchmarked at roughly 15x faster generation than autoregressive baselines on comparable hardware.

Because it's fully open source under Apache 2.0, with published weights and code, AutorunX runs it directly on its own GPU infrastructure — a Vast pod as the primary lane, RunPod as fallback — rather than calling a third-party API. That's what makes it the lowest-cost song-generation path on the platform: no per-call provider markup, just AutorunX's own compute.

On AutorunX, ACE-Step can optionally be paired with RVC (Retrieval-based Voice Conversion) for singer/voice conversion, letting a generated vocal take on a different singer's timbre after the song itself has been composed and rendered.