ACE-Step 1.5 XL in ComfyUI: comparison and remix tutorial

Compare ACE-Step 1.5 XL with earlier models in ComfyUI, test APG guidance, and explore remixes with an eight-step Turbo workflow.

Player not loading? Watch on YouTube

This tutorial compares ACE-Step 1.5 XL Turbo and SFT with their earlier counterparts in ComfyUI, then tests XL Base and several remix workflows. The speaker states that the new XL models need at least 12GB VRAM with their Gradio interface.

The speaker judges the XL examples cleaner than the older generations, especially the vocals, but notes that YouTube may obscure the difference. These are listening impressions from the demonstrated samples. For SFT and Base, he recommends adaptive projected guidance (APG) and demonstrates what happens when he disables it. His Base examples at CFG 1 with 16 steps and CFG 7 with 70 steps sound similar to him; he describes Turbo as producing good results in eight steps.

The optional lyrics section uses Ollama with Gemma 4E4B as a local LLM to generate lyrics and song tags, including prompts based on images. The speaker reports slow performance with Gemma 4 when Ollama flash attention is enabled and advises disabling it. The demonstrated nodes lack audio input.

Later experiments reuse generated latents with new conditioning, average conditioning, blend latents, and inject latent noise. The examples explore changing lyrics or musical tags while retaining elements of an earlier song, with results the speaker describes as sometimes adding vocals and harmonies.