diff --git a/renderdoc/driver/gl/gl_overlay.cpp b/renderdoc/driver/gl/gl_overlay.cpp
index cf03c03a5..7f32ac6dd 100644
--- a/renderdoc/driver/gl/gl_overlay.cpp
+++ b/renderdoc/driver/gl/gl_overlay.cpp
@@ -2162,6 +2162,18 @@ ResourceId GLReplay::RenderOverlay(ResourceId texid, FloatVector clearCol, Debug
                                      eGL_CLAMP_TO_EDGE);
 
           drv.glFramebufferTextureLayer(eGL_FRAMEBUFFER, dsAttach, overridedepth, 0, 0);
+
+          // this is a MEGA hack for mesa's behaviour - which is spec compliant. The problem is that
+          // the spec is absolutely stupid. When we called CopyTex2DMSToArray() above the
+          // framebuffer might have been incomplete since we hadn't bound this overridedepth here
+          // that we are setting up. This means that the effective stencil bits available in the FBO
+          // is 0, and mesa then clamps the query of the current stencil reference to 0. The render
+          // state push + pop inside CopyTex2DMSToArray then doesn't function correctly as it leaves
+          // the reference at 0 instead of whatever it should be - so we reset it here
+          GL.glStencilFuncSeparate(eGL_FRONT, rs.StencilFront.func, rs.StencilFront.ref,
+                                   rs.StencilFront.valuemask);
+          GL.glStencilFuncSeparate(eGL_BACK, rs.StencilBack.func, rs.StencilBack.ref,
+                                   rs.StencilBack.valuemask);
         }
 
         drv.glBindFramebuffer(eGL_DRAW_FRAMEBUFFER, curdrawfbo);