Ran Udio through its paces like a chaos monkey on a k8s cluster. 120 tracks generated, 8 actually made the cut for my project.
That's a ~6.7% success rate. For context, my typical workflow is generating stems/ideas, not finished tracks. Most outputs are like a pod stuck in `CrashLoopBackOff`—interesting failure, but not usable. The keepers? Solid foundations. Had to remix and master them heavily, though.
Is that ratio good? Depends on your tolerance for noise. If you're just prompt-bashing hoping for a hit, it's brutal. If you're using it as an idea engine within a proper GitOps-style pipeline (generate -> filter -> refine -> deploy), it's a viable, if inefficient, component.
```yaml
# My rough workflow
pipeline:
generate:
tool: udio
params: [ "prompt_variants", "style_mashes" ]
filter:
method: "manual_review"
criteria: [ "usable_hook", "clean_stems" ]
refine:
tools: [ "daw", "external_vsts" ]
deploy: # lol
target: "soundcloud_dump"
```
Your metrics will vary wildly based on prompt engineering. A bad prompt is like a misconfigured `livenessProbe`—everything dies.
Your "chaos monkey" analogy is spot on! That 6-7% success rate actually sounds pretty decent for raw idea generation. I treat Udio the same way, like a rapid prototyping tool. The real magic happens in your filter and refine stages.
One thing that changed my ratios was building a better prompt "template" for the generate phase. Instead of just mashing styles, I started defining a target BPM, a simple structure (like "intro-verse-chorus"), and even listing a couple of instruments to avoid. It cut down on the total noise significantly.
Your pipeline YAML is a vibe, by the way. I have a similar Airtable setup for tracking prompt combos vs. output quality. It feels less like random generation and more like a proper experiment then.