Closed bhouston closed 11 months ago
Render Fidelity tests were initially failing because of submodule support issues. But I got those fixed.
Now they are failing because of this difference in the Khronos-TextureTransformMulti-Test - specifically the normal rows look different:
I am investigating it now.
It looks like this change to the underlying asset makes the different renders appear to render more similar results. Thus I suggest we update the goldens for model-viewer & filament and go with this.
This is based on this analysis - that Babylon and gltf-model-viewer do not change their results, but the end results for everyone is much closer. I suspect that the way the normals were done before were exposing a bug in model-viewer and filament that was not handled but now that the asset is a bit more standard, that bug is no longer exposed.
BabylonJS's results on this test do not differ:
And filament results differ like model-viewer/three.js:
But gltf-model-viewer does not differ:
I agree with your analysis and suggestion - this is a pretty common type of update we need to do as sample bugs get fixed.
Per a previous discussion here, I have switched from the obsolete glTF-Sample-Models git repository for the various models used in the fidelity test to the newer glTF-Sample-Assets repository. I have also switched from a custom git checkout script to just using git submodules.
I had to remove a few tests from config.json unfortunately as they were relying on models that are not available in glTF-Sample-Assets because of licensing issues.