It's probably best to understand Lenstra's Elliptic Curve factorization algorithm by way of contrast with its predecessors, the Pollard's p-1 method, the Williams' p+1 method and the Cyclotomic Polynomial method of Bach and Shallit.
These are all Algebraic-group factorisation algorithms which require you to select a stage 1 bound $B_1$ and stage 2 bound $B_2$ where $B_1<B_2$.
The most common way for Pollard's p-1 method to fail is if none of the primes dividing the composite N to be factored have the order of the group of multiplicative elements (i.e p-1) able to be written as a product of prime powers smaller than $B_1$ and at most one prime between $B_1$ and $B_2$. Numbers having this representation are rare and can be loosely termed "smooth".
This is purely a characteristic of the composite to be factored and without knowledge of the factors you don't know whether the p-1 method is doomed to failure.
The Williams' p+1 method uses the multiplicative group of a quadratic extension of N which (if successfully chosen) has order p+1. So it fails if none of the $p+1$s are smooth .
The Cyclotomic Polynomial method succeeds if let's say $p^2+p+1$ or more complex expressions are smooth but these numbers are getting larger than $p$ and the chance that they are smooth (unless N has a special structure) is rapidly decreasing.
What would be nice would be to have an effectively unlimited number of different groups with different orders that we can try where there's nothing stopping the group orders being smooth and the group operations are reasonably quick. That's what Lenstra's Elliptic Curve method delivers.
The way it works is that you choose a random $\sigma$ which specifies an elliptic curve where the orders of the multiplicative groups is somewhere in the range $p+1 \pm \sqrt p$. The algorithm will succeed if for one of the curves you try, the order is smooth.
In terms a layman might understand, when Pollard's p-1 algorithm was discovered, we could all of a sudden quickly factorize a whole swathe of numbers where p-1 is smooth but we were still stuck on other numbers. When Williams' p+1 algorithm came along then we could now factorize a whole lot more numbers where p+1 is smooth but we were still stuck on lots of numbers. When Bach and Shallit's Cyclotomic Polynomial method came along, a few more numbers were factorizable but it's clear that the effectiveness of each successive algorithm was diminishing. With elliptic curves, each $\sigma$ value is essentially making available a new algorithm as good as p-1 which works if $p+1-a$ is smooth for some $|a| \lt \sqrt p$. Unlike previous algorithms, we don't know what $a$ will be until the factor has been found. Now we no longer have to worry about the prime factors being small but "hard". The difficulty is purely dependent on their size.
