We often get asked...
Did you change your models recently?
When is [next version] model coming up?
Let's tackle those one at a time!
We don't change our models every day!
While it may seem that the varying outputs you get are due to frequent model tweaks, this actually isn't the case; we're careful to do model training (and testing!) in batches. Although we unveil new and improved features at a more frequent clip, we update our models much less often, and you'll know when we do because we either give them boring version numbers or sometimes decent and helpful names.
So why do you get different results?
To quote Harriet in the song above, because "genAI is stochastic [random], yo!". That's often fun and inspiring, but admittedly sometimes also a bit... challenging, we understand!
But there are things you can do to ensure more consistency and predictability in your output when that's what you want!
Consider using manual mode. This bypasses our (genre-tag) 'prompt expansion'
From there, with identical prompts, settings, and seeds (in Advanced Settings), you'll get identical outputs*. Using that as a baseline, you can then thoughtfully and carefully make small adjustments to finetune your output more granularly.
* this is how Udio works today; given that this is an excitingly evolving space, we want to be up-front in noting that we can't guarantee that matching prompt/setting/seed input will forevermore produce identical outputs.
Okay, fine, so when's the next model coming out?
When it's ready!
Or, in a bit more detail, here's how things go:
We do extensive research in crafting a new model
We then test promising new models within our team
From there, we either send it back to the drawing board or ask our Trusted Testers to try it out across dozens of different genres.
From their input, we make a no / no-go (public release) decision.
Just know that at any given moment, we're cooking up useful & fun & very musical improvements in our models; it'll just require some patience on your end in the meantime :-)