Images used by textures with different color spaces?
donmccurdy opened this issue · 6 comments
The spec discusses image color spaces under Images, Section 3.8.3:
Effective transfer function (encoding) is defined by a glTF object that refers to the image (in most cases it’s a texture that is used by a material).
I believe this implies that the same texture object cannot be associated with multiple color spaces, and so cannot be used both as a baseColorTexture and occlusionTexture. Would it be valid for the same image object to be used by two textures in those material slots?
I've run into a question in glTF Transform, where it merges two identical textures, resulting in a color space difference in three.js (which expects 1 color space per glTF texture), and I'm not sure which is correct.
+1
Would it be valid for the same image object to be used by two textures in those material slots?
Doing this is probably unusual, but I think it is valid. @lexaknyazev What do you think? For reference, if you load the two models in the referenced glTF-Transform issue in Babylon.js sandbox, they look the same before and after.
Hello,
I got this case in a report last year:
KhronosGroup/glTF-Blender-IO#1800
Would it be valid for the same image object to be used by two textures in those material slots?
Doing this is probably unusual, but I think it is valid. @lexaknyazev What do you think? For reference, if you load the two models in the referenced glTF-Transform issue in Babylon.js sandbox, they look the same before and after.
I just checked Babylon viewer and indeed they look the same,
I inspected with spector, babylon viewer isn't using sRGB color spacing which was the problem originally, this may be why they both look the same on babylon
babylon viewer isn't using sRGB color spacing which was the problem originally
I think you mean the sandbox doesn't use EXT_sRGB formats. If so, yes, by default, we have this turned off because of a Chromium issue. If you use https://3dcommerce.babylonjs.com, it will use EXT_sRGB formats. The result is the same though.
FWIW, three.js has moved to use texStorage2D which avoids the Chromium performance issues. We do still see similar performance issues on video textures, and use inline sRGB decoding in the shader to handle that instead.
While investigating support for wide gamut colorspaces in WebGL and three.js, I've encountered an issue that might be relevant here. Images are unpacked by WebGL into a target color space identified by gl.unpackColorSpace. The web browser infers the source color space of an image from its ICC profile, and uses that to make the required conversion into the unpack color space. While the glTF Spec advises setting the "UNPACK_COLORSPACE_CONVERSION_WEBGL flag to NONE," we can no longer do that in scenes whose textures use more than one set of color primaries, or where image primaries do not match the primaries of the working/rendering color space, without re-implementing the entire image decoding pipeline ourselves (and we aren't going to do that).
As a result — I think the glTF specification's current advice about ignoring ICC Profiles is still reasonable, however, I do not believe we should be encouraging situations where the embedded ICC Profile cannot possibly be accurate, as would be the case when the image contains mixed data from different color spaces.
EDIT: Perhaps this is something the validator should try to warn about?