I stumbled upon this paper: http://graphics.cs.williams.edu/papers/CSSM/

Here's the video:

It looks pretty amazing.

Has anyone else tried something like this? I'm thinking of trying it out in my engine, but if anyone knows that it creates some pretty bad looking artifacts I would probably avoid it.

I don't think it looks that good to be honest because it is not close enough to how real caustics behave. I would probably prefer ordinary shadow maps instead.

Edit: I'm referring to the video. Some of the examples in the paper look a bit more convincing.

I just noticed at about :15 seconds in the shadows inside the little tube things didn't look too good. I think the purple shadows from the purple pipes and blueish shadow from the fence look really nice, although I'm really confused how that red shadow in the beginning is cast by that purple tank.

I remember seeing this a few years ago and dismissing it as being too expensive (2-4x memory use, and larger filters required)... but my current project has a lot of layers of translucency, and GPUs can now deal with the bandwidth requirements for this... so I might give it a go and see how it looks

I don't think it looks that good to be honest because it is not close enough to how real caustics behave.

It doesn't model caustics at all, so anything that looks like caustics is just an artefact of the filtering

Ordinary shadow maps are an apples-and-oranges comparison, because they don't support translucency at all.

Well, it is caustics (light passing through a transparent substance) without taking refraction into account. For paper thin objects it would be a close approximation, but for those giant colored tubes it fails completely and is IMO more distracting than an ordinary shadow.