These are not really parallel technologies, but simplistically:
When you are using the SAML model, you typically respect a single issuer of SAML tokens (or perhaps a handful, but pre-defined by the service). That is, there is a single, central repository of user credentials, and all the "consumers" accept its' tokens and validate against that one.
With OpenId, you are typically accepting any OpenId provider, based on the user's input - it could be Google, or Facebook, or StackExchange, or your corporate OpenId provider, or your own home server. Either way, it doesn't matter.
This is not just a question of numbers, either. With SAML, you explicitly trust the SAML server to authenticate and verify your actual identity - which you can later rely on. With OpenId, you don't necessarily have any trust relationship with the OpenId provider - you accept whatever it tells you, and you don't actually care: you just need some form of assured identity (actually more of a "placeholder" in this case).
So, to sum up: SAML is centralized because you expect to control identities, OpenId is distributed because you have given up that control.