NormalMap: Realtime vs Vray

[rasmus]

Dear chaotic community,

I've recently come across a specific problem, which I hope somebody can help me solve:

I've received data from a client (obj + normal map). I assume they scanned it via photogrammetry and did some kind of poly reduction. To verify the results the client is testing geometry with the normal map in adobe painter and sketchfab. In both dases the results are good, except for a few small areas which I don't need worry about (marked green).

However the same combination of obj + normal map give me much more broken results in max2022 and vray5, which I can not explain. By the way, I know that the geo on its own looks terrible and so does the normalmap, but I've seen this very often after poly reduction so I know that bad looking normal maps combined with bad looking topology can achieve perfect realtime results).

The settings for the material setup are extremely simple, however I might have missed something? Here's the setup:
The vray bitmap with the normalmap (linear color space, gamma 1.0) is plugged into vray NormalBump node feeding the bump slot of the vray shader.
The scene is lit with a simple HDR, nothing fancy. All render settings are at default values (Color mapping only, no gamma) and so is the vray material and the vray normal bump node. The resulting render looks like the attached image. Its easy to see that there are major problems on the top edge and the bottom as well. I've tried flipping the green channel, but this doesn't really change anything. I even tried flipping the green channel in PS manually in case its a bug but this won't yield better results. Out of curiosity and desperation I've tested all possible combinations of flipping without success.

All of this leads to the following question:

The standard bump value in max is set to 30, and the standard value for vray normal maps is set to 1. Are those valid values, or should max bump intensity be set to 100 for flawless normal map rendering? In the past I've always understood those values as "artistic choices" so I tried increasing and decreasing both bump intensity values (vray normal mal value and vray shader bump intensity) but no combination gave me the same good results of the realtime version (judging with a running progessive preview by the way).

What other option could be causing this strange behaviour of all realtime applications creating better results than vray? Please keep in mind, that "fixing" the geometry normals should not be part of the process simply because it isn't neccessary in realtime applications, and again, I've seen weird looking geo with weird normal maps create perfect results very often.

I will gladly supply the obj and the normalmap for everyone to test this. Any help is highly appreciated! =)

Best regards,
Rasmus

[rasmus]

Due to filesize restrictions in the forum I've uploaded the obj and the normalmap to sendspace:
https://www.sendspace.com/filegroup/...hy%2Fh%2Fev7TA

[hermit.crab]

Try flipping green and increasing the shader's bump map value to 100. This produces similar results to the reference.

[rasmus]

Well, thanks for the quick reply. I did that and I don't think the results are even close. Could you please post a screenshot of your rendering so we have a visual reference of what you consider to be "similar"?

[kosso_olli]

There was a problem discussed recently looking very similar to yours: https://forums.chaosgroup.com/forum/...aps-in-v-ray-5

[rasmus]

Thank you very much for the link. I had searched prior to posting but didn't see this thread. It seems to me that max/vray have some kind of bug at this point, which is obvious if other render engines dont produce the same mistakes using the same meshes and maps.

I'll gladly follow the change logs and updates on this post and on the one you linked above.

[^Lele^]

The issue you have seems to not be the same as the one linked: in your case, both Corona and V-ray produce identical results.

The issue stems from the fact that you're using object-space (or world-space) normal maps, and V-Ray only supports tangent-space normal maps.

You can fix this in painter via loading and conversion to tangent space, i'd assume.

This can be verified in Houdini, which with tangent-space mapping produces results identical to V-Ray and Corona, but crucially has a way to switch the interpretation of the normal map's vector space.
I'll pass the file setup to the relevant Dev for inspection.