ACE-Step 1.5 XL

Open diffusion-based music-generation foundation model by the ACE-Step team — 3.5B parameters, transformer-conditioned, generating music from text rather than speech. Fifteen cards carry a recipe here and the split is clean: the six 12 GB cards go through the project's 8 GB optimisation path, while everything from 16 GB up runs it straight in ComfyUI, both Radeon boards included. Apache-2.0.