TITLE: GIT and singularities
QUESTION [7 upvotes]: Let $G$ be a complex reductive group acting on a complex affine variety $X$ and let $X // G = \operatorname{Spec}\mathbb{C}[X]^G$ be the GIT quotient.
Is there a relationship between the singular locus of $X$ and that of $X // G$?
Of course, $X//G$ can be highly singular while $X$ is smooth. But, for example, I was wondering if (or under what conditions) singular points of $X$ are mapped to singular points of $X // G$.
Edit. Spenser's nice comment below shows that the answer to the latter is no. But perhaps a better and more precise question is: If $X // G$ is non-singular at $y$, is there a non-singular $x \in X$ mapping to $y$? In other words, do all fibres of $X \to X // G$ at non-singular points contain a non-singular point of $X$? I'm willing to assume irreducibility or other nice properties.

REPLY [5 votes]: This is an answer to the revised question.  It is the simplest counterexample that I can think of where the reductive group is smooth and connected, where $X$ is normal and affine, and where $Y=X//G$ is smooth, even though there are fibers of the quotient map that are contained in the singular locus of $X$.
Let $Y$ be $\text{Spec}\ k[x,y,z]$, i.e., affine $3$-space.  Let $G$ be the multiplicative group of units, $G=\text{Spec}\ k[u,u^{-1}]$.  Let $X$ be $\text{Spec}\ k[x,y,z,s,t,t^{-1}]/\langle f \rangle$ where $f$ is the polynomial, $$f=s^2+t(xz-y^2).$$  Let the action of $G$ on $X$ be defined by $$\mu:G\times_{\text{Spec}\ k} X \to X, \ \ \mu(u,(x,y,z,s,t)) = (x,y,z,us,u^2t). $$  The ring of $G$-invariant polynomials is the subring, $$k[X]^G = k[x,y,z].$$  The quotient map is just the usual projection, $$q:X\to Y, \ \ q(x,y,z,s,t) = (x,y,z).$$  For the dense Zariski open $U = D(xz-y^2)\subset Y$, the inverse image $q^{-1}(U)$ is a $G$-torsor over $U$.
The singular locus of $X$ is the single $q$-fiber, $q^{-1}(0,0,0)$.  Even though the origin is a smooth point of $Y$, every point of this $q$-fiber is a singular point of $X$.