Descriptions:
Nerdy Rodent demonstrates how to run the new ACE-Step 1.5 XL models for AI music generation locally through a custom ComfyUI workflow. The XL variants are roughly twice the size of their predecessors, require at least 12GB of VRAM, and deliver noticeably cleaner audio — particularly in vocal clarity — compared to the earlier turbo and SFT model versions.
The walkthrough covers a purpose-built ComfyUI workflow using the presenter’s color-coded ‘rodent method’ for readability. The most important component highlighted is the Adaptive Projected Guidance (APG) node, which the video demonstrates has a dramatic effect on output quality — toggling it off produces significantly degraded results regardless of CFG or step count. The workflow also integrates Ollama running Gemma 4 4B for AI-assisted lyric and tag generation, with toggle switches throughout allowing users to mix full AI generation, manual lyrics, or hybrid approaches. A practical note: Gemma 4 requires Flash Attention to be disabled in Ollama or generation becomes extremely slow.
Side-by-side audio comparisons walk through XL vs. standard models, base vs. SFT variants, and different CFG/step configurations, with the finding that the APG node matters far more than step count. An experimental section explores a conditioning-average remix technique using a second seed and different key scale. Workflow files are available via the channel’s Patreon for those wanting a ready-made starting point.
📺 Source: Nerdy Rodent · Published April 11, 2026
🏷️ Format: Tutorial Demo
