ACE-Step 1.5 XL = Free Music Generation in ComfyUI!

Summary

Key Points

• The video introduces new A-step XL models, which are larger and require at least 12GB of VRAM, and demonstrates their use within Comfy UI workflows. 0:10
• The workflow utilizes the "rodent method" of organizing nodes into colorful groups for easier understanding and updates, encouraging users to copy elements for their own projects. 0:29
• A crucial addition to A-step workflows is the Adaptive Projected Guidance (APG) node, which significantly improves audio quality, especially with SFT or base models, and can be easily toggled with a bypass switch. 1:26
• The workflow integrates AI prompting using LLMs like Ollama with Gemma 4E4B to generate or tag song lyrics and tags, noting to disable flash attention for Gemma 4 to avoid slowdowns. 2:24
• Comparative tests demonstrate that the new XL models produce slightly higher quality and cleaner audio compared to previous turbo and SFT models, though personal testing is encouraged. 5:35
• The APG node is shown to be essential for good audio quality, as disabling it results in significantly degraded output, even with adjusted temperature and top P settings. 7:46
• With the APG node enabled, users can achieve good results with fewer CFG steps (e.g., around 20 steps instead of 70), leading to quicker song generation. 10:41
• The video explores advanced experimentation techniques, including generating a second set of conditioning with different settings and using latent blending, resamplers, and conditioning average nodes to create varied remixes. 11:17
• The resampler often adds interesting variations like extra vocals and harmonies, while the conditioning average node tends to make the first set of conditionings dominate. 11:56
• The workflow allows for creative mixing of AI-generated content (lyrics, tags, melody) with user-defined elements, demonstrating how to blend different musical styles and lyrical themes. 18:38