I found that merging Stable Audio 3 Medium base and non-base at 2:1 ratio allows one to increase CFG without the model breaking down. This virtually eliminates the infamous "SA3 crunch" and considerably improves audio quality.
This merge has been converted to gguf for use in sa3.cpp
Example (16steps, cfg3)
<audio controls src="https://cdn-uploads.huggingface.co/production/uploads/66edc6ec77590e90ba7c109c/0KOsYGj8GaaPKvwJ8ThFV.mpga"></audio>
