Precompiled binaries releases for Windows, Mac OS X and Linux may be found here.
The tool can be invoked like so:
> FBX2glTF ~/models/butterfly.fbx
Or perhaps, as part of a more complex pipeline:
> FBX2glTF --binary --draco --verbose \ --input ~/models/source/butterfly.fbx \ --output ~/models/target/butterfly.glb
You can always run the binary with --help to see what options it takes:
FBX2glTF 2.0: Generate a glTF 2.0 representation of an FBX model. Usage: FBX2glTF [OPTION...] [<FBX File>] -i, --input arg The FBX model to convert. -o, --output arg Where to generate the output, without suffix. -e, --embed Inline buffers as data:// URIs within generated non-binary glTF. -b, --binary Output a single binary format .glb file. -d, --draco Apply Draco mesh compression to geometries. --flip-u Flip all U texture coordinates. --flip-v Flip all V texture coordinates (default behaviour!) --no-flip-v Suppress the default flipping of V texture coordinates --pbr-metallic-roughness Try to glean glTF 2.0 native PBR attributes from the FBX. --khr-materials-unlit Use KHR_materials_unlit extension to specify Unlit shader. --blend-shape-normals Include blend shape normals, if reported present by the FBX SDK. --blend-shape-tangents Include blend shape tangents, if reported present by the FBX SDK. --long-indices arg Whether to use 32-bit indices (never|auto|always). --compute-normals arg When to compute normals for vertices (never|broken|missing|always). -k, --keep-attribute arg Used repeatedly to build a limiting set of vertex attributes to keep. -v, --verbose Enable verbose output. -h, --help Show this help. -V, --version Display the current program version.
Some of these switches are not obvious:
--embedis the way to get a single distributable file without using the binary format. It encodes the binary buffer(s) as a single enormous base64-encoded
data://URI. This is a very slow and space-consuming way to accomplish what the binary format was invented to do simply and efficiently, but it can be useful e.g. for loaders that don't understand the .glb format.
--flip-v, when enabled, will apply a
x -> (1.0 - x)function to all
vtexture coordinates respectively. The
uversion is perhaps not commonly used, but flipping
vis the default behaviour. Your FBX is likely constructed with the assumption that
(0, 0)is bottom left, whereas glTF has
(0, 0)as top left. To produce spec-compliant glTF, we must flip the texcoords. To request unflipped coordinates:
--long-indiceslets you force the use of either 16-bit or 32-bit indices. The default option is auto, which make the choice on a per-mesh-size basis.
--compute-normalscontrols when automatic vertex normals should be computed from the mesh. By default, empty normals (which are forbidden by glTF) are replaced. A choice of 'missing' implies 'broken', but additionally creates normals for models that lack them completely.
--no-flip-vwill actively disable v coordinat flipping. This can be useful if your textures are pre-flipped, or if for some other reason you were already in a glTF-centric texture coordinate system.
- All three material options are, in their own way, works in progress, but the
--pbr-metallic-roughnessswitch is at least compliant with the core spec; unlike the others, it does not depend on an unratified extension. That option will be chosen by default if you supply none of the others. Material switches are documented further below.
- If you supply any
-keep-attributeoption, you enable a mode wherein you must supply it repeatedly to list all the vertex attributes you wish to keep in the conversion process. This is a way to trim the size of the resulting glTF if you know the FBX contains superfluous attributes. The supported arguments are
- When blend shapes are present, you may use
--blend-shape-tangentsto include normal and tangent attributes in the glTF morph targets. They are not included by default because they rarely or never seem to be correctly present in the actual FBX source, which means the SDK must be computing them from geometry, unasked? In any case, they are beyond the control of the artist, and can yield strange crinkly behaviour. Since they also take up significant space in the output file, we made them opt-in.
Building it on your own
This build process has been tested on Linux, Mac OS X and Windows. It requires CMake 3.5+ and a reasonably C++11 compliant toolchain.
You must manually download and install the Autodesk FBX SDK and accept its license agreement.
At present, only version 2018.1.1 of the FBX SDK is supported. The build system will not successfully locate any other version.
Linux and MacOS X
Compilation on Unix machines should be as simple as:
> cd <FBX2glTF directory> > cmake -H. -Bbuild -DCMAKE_BUILD_TYPE=Release > make -Cbuild -j4 install
If all goes well, you will end up with a statically linked executable.
As part of this process, you will be asked to choose which generator to use. At present, only Visual Studio 2017 is supported. Older versions of the IDE are unlikely to successfully build the tool.
(MinGW support is plausible. The difficulty is linking statically against the FBX SDK .lib file. Contributions welcome.)
Note that the
CMAKE_BUILD_TYPE variable from the Unix Makefile system is
entirely ignored here; it is when you open the generated solution that
you will be choose one of the canonical build types — Debug,
Release, MinSizeRel, and so on.
The actual translation begins with the FBX SDK parsing the input file, and ends
with the generation of the descriptive
JSON that forms the core of glTF, along
with binary buffers that hold geometry and animations (and optionally also
emedded resources such as textures.)
In the process, each mesh is ripped apart into a long list of triangles and their associated vertices, with a material assigned to each one. A similar process happens in reverse when we construct meshes and materials that conform to the expectations of the glTF format.
Every animation in the FBX file becomes an animation in the glTF file. The method used is one of "baking": we step through the interval of time spanned by the animation, keyframe by keyframe, calculate the local transform of each node, and whenever we find any node that's rotated, translated or scaled, we record that fact in the output.
Beyond skeleton-based animation, Blend Shapes are also supported; they are read from the FBX file on a per-mesh basis, and clips can use them by varying the weights associated with each one.
The baking method has the benefit of being simple and precise. It has the drawback of creating potentially very large files. The more complex the animation rig, the less avoidable this data explosion is.
There are three future enhancements we hope to see for animations:
- Version 2.0 of glTF brought us support for expressing quadratic animation curves, where previously we had only had linear. Not coincidentally, quadratic splines are one of the key ways animations are expressed inside the FBX. When we find such a curve, it would be more efficient to output it without baking it into a long sequence of linear approximations.
- We do not yet ever generate sparse accessors, but many animations (especially morph targets) would benefit from this storage optimisation.
- Perhaps most useful in practice is the idea of compressing animation curves the same way we use Draco to compress meshes (see below). Like geometry, animations are highly redundant — each new value is highly predictable from preceding values. If Draco extends its support for animations (it's on their roadmap), or if someone else develops a glTF extension for animation compression, we will likely add support in this tool.
With glTF 2.0, we leaped headlong into physically-based rendering (PBR), where the canonical way of expressing what a mesh looks like is by describing its visible material in fundamental attributes like "how rough is this surface".
By contrast, FBX's material support remains largely in the older world of Lambert and Phong, with simpler and more direct illumination and shading models. These modes are inherently incompatible — for example, textures in the old workflow often contain baked lighting of the type that would arise naturally in a PBR environment.
Some material settings remain well supported and transfer automatically:
- Emissive constants and textures
- Occlusion maps
- Normal maps
This leaves the other traditional settings, first of Lambert:
- Ambient — this is anathema in the PBR world, where such effects should emerge naturally from the fundamental colour of the material and any ambient lighting present.
- Diffuse — the material's direction-agnostic, non-specular reflection, and additionally, with Blinn/Phong:
- Specular — a more polished material's direction-sensitive reflection,
- Shininess — just how polished the material is; a higher value here yields a more mirror-like surface.
(All these can be either constants or textures.)
Exporting as Unlit
If you have a model was constructed using an unlit workflow, e.g. a photogrammetry capture or a landscape with careful baked-in lighting, you may choose to export it using the --khr-materials-common switch. This incurs a dependency on the glTF extension 'KHR_materials_unlit; a client that accepts that extension is making a promise it'll do its best to render pixel values without lighting calculations.
Note that at the time of writing, this glTF extension is still undergoing the ratification process
Exporting as Metallic-Roughness PBR
Given the command line flag --pbr-metallic-roughness, we throw ourselves into the warm embrace of glTF 2.0's PBR preference.
As mentioned above, there is little consensus in the world on how PBR should be represented in FBX. At present, we support only one format: Stingray PBS. This is a feature that comes bundled with Maya, and any PBR model exported through that route should be digested propertly by FBX2glTF.
(A happy note: Allegorithmic's Substance Painter also exports Stingray PBS, when hooked up to Maya.)
The tool will optionally apply Draco compression to the geometric data of each mesh (vertex indices, positions, normals, per-vertex color, and so on). This can be dramatically effective in reducing the size of the output file, especially for static models.
Enabling this feature adds an expressed required dependency in the glTF on the
KHR_draco_geometry_compression extension, and can thus only be loaded by a
viewer that is willing and able to decompress the data.
Note that at the time of writing, this glTF extension is still undergoing the ratification process.
This tool is under continuous development. We do not have a development roadmap per se, but some aspirations have been noted above. The canonical list of active TODO items can be found on GitHub.
- Pär Winzell
- J.M.P. van Waveren
- Amanda Watson
FBX2glTF is BSD-licensed. We also provide an additional patent grant.