Able to reproduce this issue on reported version 66.0.3359.106 and on latest canary 68.0.3397.0 using Windows 10, Mac 10.13.3 and Ubuntu 14.04.

This issue is seen from M-60. Hence considering this issue as Non-Regression and marking as Untriaged.

NOTE:Issue is working fine in Firefox

Thanks!

For some reason when it gets a composited layer (will-change: transform is enough) the image sampling goes bad. vmpstr@, do you know why this is?

When the image is promoted to a separate layer, it goes down the directly composited image path, which rasterizes at a power of two scale and then uses the the texture throughout. We also have scale banding to prevent changing the scale frequently on up-and-down animations. All in all, in this specific example, we rasterize the image at scale 1 (height: 768) and then use the gpu to downscale it to the needed scale (height: 300). In the non-directly composited case, we downscale the image to a power of two (height: 384), rasterize it, and use the gpu to scale down to needed scale (height: 300). 

Now I've tried to hack the code to rasterize at scale 0.5 initially, which theoretically should produce the same results, but it didn't. It did look different, but not the same. 

As for why the image is being directly composited, I imagine the intent is that if you specify a backface visibility, it's a hint that a transform is likely to be applied which would promote this to a layer anyway. 

-> khushalsagar to see if anything can be done here, although my feeling is that this is one of the cases where we can't do much

+ericrk as well.