I have just confirmed it on Windows as well, with newer drivers than yours @Philipp Oeser (lichtwerk)

I can't do that right now. I did however confirm it on Windows, with the latest NVidia drivers.

Thanks for looking into this! I've also occasionally seen those vanishing faces when it didn't crash. Also, wasn't aware of the update_data argument, either.

I'm opening this one again, because the issue seems to have re-appeared in rBcb7ead2e3b62. Seems like it's - in my case at least - for 3+ Materials now, so make sure to use the cube on the right in the example blend above.

Looks like there are similar, and more issues in 2.79:

active.data = bpy.data.meshes.new_from_object(active.evaluated_get(dg), preserve_all_data_layers=True, depsgraph=dg)

Sorry for the delay, just wanted to confirm, that this is working now. Thanks @Sergey Sharybin (sergey)!

Thank you, I've tried invalidate_caches(), but it didn't seem to have any affect. reload() works very well! Thanks!

Yep, thank you!

In an attempt to avoid it being overlooked, I want to point out, that it's not just Subsurface, but also Clear Coat and Transmission causing this. Also, as you can see in the file attached, there is an issue with the Alpha input as well, which leads to darkening of the material.

It is also not clear what the purpose is of using alpha twice in the shader node setup, as using the Alpha socket means it's already mixing with a Transparent BSDF internally.

My use case is this:
I setup a modular object using 1 base object with one array mods, and 2 cap objects. I prepare the base object and the caps with vertex groups and I want the vertex groups to be merged on the evaluated mesh.
This works perfectly when using the Apply button in the modifier panel, but not using bpy.data.meshes.new_from_object()
To be more specific, here's is how I use this: https://youtu.be/NyZfF8HtTUU?t=669. I use the vertex group to mark specific faces, which need their normals to be adjusted, among other things, after the mods have been applied and the object is joined with another one and integrated into the surface.

Like https://developer.blender.org/T64300, this has now also been fixed by rBbe5fd9c09f25. Thank you @Sergey Sharybin (sergey)! Closing.

Actually, changing the color space to linear on a single texture node also seems to work.

That is interesting Will, thanks for pointing that out. I wonder why this happens.

I'm still on the old build. I assume not, if this is the fix: mesh_in_bmain->smoothresh = mesh->smoothresh;

Awesome, wasn't sure if it's a bug or intended behavior. And while at it. It also doesn't/didn't preserve use_auto_smooth and I assume auto_smooth_angle

The issues persists, even after the recent depsgraph changes.
I have update the blend:

@Sergey Sharybin (sergey) Thanks, you've been very helpful. How would you apply mods now, while keeping the materials assigned as they are by mods like bevel for instance. See

I suspect rB7ad802cf3ae5 is the reason.
@Brecht Van Lommel (brecht) Is there any solution, that doesn't involve having to unpack the alpha channel in a separate image?

Happy to help. Good luck.

Check out was Sergey told me:  https://developer.blender.org/T64735#681264

Have you tried bpy.data.meshes.new_from_obj(obj_eval)?

Using active.data = bpy.data.meshes.new_from_object(active_eval)
as mentioend here, seems to work.

What build, what OS, what Graphics card?

Confirmd.

It may be related, but I'm getting a clear error message from this one here, related to the binding/unbinding of the deformation mods, whereas I can only observe the misbehavior in the other one (and it misbehaves in a different way as well).
I think that merits a separate bug report.

If I throw in a Mesh Deform mod as well, there will be a new error (bind, instead of unbind)

FWIW, if you rotate an object and then parent it, you will have the same problem. The object won't be parented in it's new rotated orientation, as you'd expect.

My issue seems to have been resolved in the lasted build bot rB480a09a92f7f

Looks to me like it has been fixed by rBd2f1705b15be. Closing.

Is the parsing of the commit hash link broken? Noticed it yesterday already.