Configure the model conversion

This chapter documents the options for the model conversion.

Settings file

If a file called <modelName>.ConversionSettings.json is found in the input container beside the input model <modelName>.<ext>, then it will be used to provide additional configuration for the model conversion process. For example, box.ConversionSettings.json would be used when converting box.gltf.

The contents of the file should satisfy the following json schema:

    "$schema" : "",
    "description" : "ARR ConversionSettings Schema",
    "type" : "object",
    "properties" :
        "scaling" : { "type" : "number", "exclusiveMinimum" : 0, "default" : 1.0 },
        "recenterToOrigin" : { "type" : "boolean", "default" : false },
        "opaqueMaterialDefaultSidedness" : { "type" : "string", "enum" : [ "SingleSided", "DoubleSided" ], "default" : "DoubleSided" },
        "gammaToLinearMaterial" : { "type" : "boolean", "default" : false },
        "gammaToLinearVertex" : { "type" : "boolean", "default" : false },
        "sceneGraphMode": { "type" : "string", "enum" : [ "none", "static", "dynamic" ], "default" : "dynamic" },
        "generateCollisionMesh" : { "type" : "boolean", "default" : true },
        "unlitMaterials" : { "type" : "boolean", "default" : false },
        "fbxAssumeMetallic" : { "type" : "boolean", "default" : true },
        "deduplicateMaterials" : { "type" : "boolean", "default" : true },
        "axis" : {
            "type" : "array",
            "items" : {
                "type" : "string",
                "enum" : ["default", "+x", "-x", "+y", "-y", "+z", "-z"]
            "minItems": 3,
            "maxItems": 3
        "metadataKeys": {
            "type": "array",
            "items": {
                "type": "string"
    "additionalProperties" : false

An example file box.ConversionSettings.json might be:

    "scaling" : 0.01,
    "recenterToOrigin" : true,
    "material-override" : "box_materials_override.json"

Geometry parameters

  • scaling - This parameter scales a model uniformly. Scaling can be used to grow or shrink a model, for example to display a building model on a table top. Scaling is also important when a model is defined in units other than meters, since the rendering engine expects meters. For example, if a model is defined in centimeters, then applying a scale of 0.01 should render the model at the correct size. Some source data formats (for example .fbx) provide a unit scaling hint, in which case the conversion implicitly scales the model to meter units. The implicit scaling provided by the source format will be applied on top of the scaling parameter. The final scaling factor is applied to the geometry vertices and the local transforms of the scene graph nodes. The scaling for the root entity's transform remains unmodified.

  • recenterToOrigin - States that a model should be converted so that its bounding box is centered at the origin. If a source model is displaced far from the origin, floating point precision issues may cause rendering artifacts. Centering the model can help in this situation.

  • opaqueMaterialDefaultSidedness - The rendering engine assumes that opaque materials are double-sided. if that assumption isn't true of a particular model, this parameter should be set to "SingleSided". For more information, see single sided rendering.

Material de-duplication

  • deduplicateMaterials - This parameter enables or disables automatic de-duplication of materials that share the same properties and textures. De-duplication happens after material overrides have been processed. It's enabled by default.

  • If even after de-duplication a model has more than 65,535 materials, the service will attempt to merge materials with similar properties. As a last resort any materials exceeding the limit will be replaced by a red error material.

Image shows two cubes of 68,921 colored triangles.

Two cubes of 68,921 colored triangles. Left: Before de-duplication with 68,921 color materials. Right: After de-duplication with 64,000 color materials. The limit is 65,535 materials. (See limits.)

Color space parameters

The rendering engine expects color values to be in linear space. If a model is defined using gamma space, then these options should be set to true.

  • gammaToLinearMaterial - Convert material colors from gamma space to linear space
  • gammaToLinearVertex - Convert vertex colors from gamma space to linear space


For FBX files these settings are set to true by default. For all other file types, the default is false.

Scene parameters

  • sceneGraphMode - Defines how the scene graph in the source file is converted:
    • dynamic (default): All objects in the file are exposed as entities in the API and can be transformed independently. The node hierarchy at runtime is identical to the structure in the source file.
    • static: All objects are exposed in the API but they cannot be transformed independently.
    • none: The scene graph is collapsed into one object.

Each mode has different runtime performance. In dynamic mode, the performance cost scales linearly with the number of entities in the graph, even when no part is moved. Use dynamic mode only when it is necessary to move parts individually, for example for an 'explosion view' animation.

The static mode exports the full scene graph, but parts inside this graph have a constant transform relative to its root part. The root node of the object, however, can still be moved, rotated, or scaled at no significant performance cost. Furthermore, spatial queries will return individual parts and each part can be modified through state overrides. With this mode, the runtime overhead per object is negligible. It is ideal for large scenes where you still need per-object inspection but no per-object transform changes.

The none mode has the least runtime overhead and also slightly better loading times. Inspection or transform of single objects is not possible in this mode. Use cases are, for example, photogrammetry models that do not have a meaningful scene graph in the first place.


Many applications will load multiple models. You should optimize the conversion parameters for each model depending on how it will be used. For example, if you want to display the model of a car for the user to take apart and inspect in detail, you need to convert it with dynamic mode. However, if you additionally want to place the car in a show room environment, that model can be converted with sceneGraphMode set to static or even none.

Physics parameters

  • generateCollisionMesh - If you need support for spatial queries on a model, this option has to be enabled. In the worst case, the creation of a collision mesh can double the conversion time. Models with collision meshes take longer to load and when using a dynamic scene graph, they also have a higher runtime performance overhead. For overall optimal performance, you should disable this option on all models on which you don't need spatial queries.

Unlit materials

  • unlitMaterials - By default the conversion will prefer to create PBR materials. This option tells the converter to treat all materials as color materials instead. If you have data that already incorporates lighting, such as models created through photogrammetry, this option allows you to quickly enforce the correct conversion for all materials, without the need to override each material individually.

Converting from older FBX formats, with a Phong material model

  • fbxAssumeMetallic - Older versions of the FBX format define their materials using a Phong material model. The conversion process has to infer how these materials map to the renderer's PBR model. Usually this works well, but an ambiguity can arise when a material has no textures, high specular values, and a non-grey albedo color. In this circumstance, the conversion has to choose between prioritizing the high specular values, defining a highly reflective, metallic material where the albedo color dissolves away, or prioritizing the albedo color, defining something like a shiny colorful plastic. By default, the conversion process assumes that highly specular values imply a metallic material in cases where ambiguity applies. This parameter can be set to false to switch to the opposite.

Coordinate system overriding

  • axis - To override coordinate system unit-vectors. Default values are ["+x", "+y", "+z"]. In theory, the FBX format has a header where those vectors are defined and the conversion uses that information to transform the scene. The glTF format also defines a fixed coordinate system. In practice, some assets either have incorrect information in their header or were saved with a different coordinate system convention. This option allows you to override the coordinate system to compensate. For example: "axis" : ["+x", "+z", "-y"] will exchange the Z-axis and the Y-axis and keep coordinate system handedness by inverting the Y-axis direction.

Node meta data

  • metadataKeys - Allows you to specify keys of node metadata properties that you want to keep in the conversion result. You can specify exact keys or wildcard keys. Wildcard keys are of the format "ABC*" and match any key that starts with "ABC". Supported metadata value types are bool, int, float, and string.

    For GLTF files this data comes from the extras object on nodes. For FBX files this data comes from the Properties70 data on Model nodes. Please consult the documentation of your 3D Asset Tool for further details.

Vertex format

It is possible to adjust the vertex format for a mesh, to trade precision for memory savings. A lower memory footprint allows you to load larger models or achieve better performance. However, depending on your data, the wrong format can significantly impact rendering quality.


Changing the vertex format should be a last resort when models don't fit into memory anymore, or when optimizing for the best possible performance. Changes can easily introduce rendering artifacts, both obvious ones and subtle ones. Unless you know what to look out for, you should not change the default.

These adjustments are possible:

  • Specific data streams can be explicitly included or excluded.
  • The accuracy of data streams can be decreased to reduce the memory footprint.

The following vertex section in the .json file is optional. For each portion that is not explicitly specified, the conversion service falls back to its default setting.

    "vertex" : {
        "position"  : "32_32_32_FLOAT",
        "color0"    : "NONE",
        "color1"    : "NONE",
        "normal"    : "NONE",
        "tangent"   : "NONE",
        "binormal"  : "NONE",
        "texcoord0" : "32_32_FLOAT",
        "texcoord1" : "NONE"

By forcing a component to NONE, it is guaranteed that the output mesh does not have the respective stream.

Component formats per vertex stream

These formats are allowed for the respective components:

Vertex component Supported formats (bold = default)
position 32_32_32_FLOAT, 16_16_16_16_FLOAT
normal 8_8_8_8_SIGNED_NORMALIZED, 16_16_16_16_FLOAT, NONE
tangent 8_8_8_8_SIGNED_NORMALIZED, 16_16_16_16_FLOAT, NONE
binormal 8_8_8_8_SIGNED_NORMALIZED, 16_16_16_16_FLOAT, NONE
texcoord0 32_32_FLOAT, 16_16_FLOAT, NONE
texcoord1 32_32_FLOAT, 16_16_FLOAT, NONE

Supported component formats

The memory footprints of the formats are as follows:

Format Description Bytes per vertex
32_32_FLOAT two-component full floating point precision 8
16_16_FLOAT two-component half floating point precision 4
32_32_32_FLOAT three-component full floating point precision 12
16_16_16_16_FLOAT four-component half floating point precision 8
8_8_8_8_UNSIGNED_NORMALIZED four-component byte, normalized to [0; 1] range 4
8_8_8_8_SIGNED_NORMALIZED four-component byte, normalized to [-1; 1] range 4

Best practices for component format changes

  • position: It is rare that reduced accuracy is sufficient. 16_16_16_16_FLOAT introduces noticeable quantization artifacts, even for small models.
  • normal, tangent, binormal: Typically these values are changed together. Unless there are noticeable lighting artifacts that result from normal quantization, there is no reason to increase their accuracy. In some cases, though, these components can be set to NONE:
    • normal, tangent, and binormal are only needed when at least one material in the model should be lit. In ARR this is the case when a PBR material is used on the model at any time.
    • tangent and binormal are only needed when any of the lit materials uses a normal map texture.
  • texcoord0, texcoord1 : Texture coordinates can use reduced accuracy (16_16_FLOAT) when their values stay in the [0; 1] range and when the addressed textures have a maximum size of 2048 x 2048 pixels. If those limits are exceeded, the quality of texture mapping will suffer.


Assume you have a photogrammetry model, which has lighting baked into the textures. All that is needed to render the model are vertex positions and texture coordinates.

By default the converter has to assume that you may want to use PBR materials on a model at some time, so it will generate normal, tangent, and binormal data for you. Consequently, the per vertex memory usage is position (12 bytes) + texcoord0 (8 bytes) + normal (4 bytes) + tangent (4 bytes) + binormal (4 byte) = 32 bytes. Larger models of this type can easily have many millions of vertices resulting in models that can take up multiple gigabytes of memory. Such large amounts of data will affect performance and you may even run out of memory.

Knowing that you never need dynamic lighting on the model, and knowing that all texture coordinates are in [0; 1] range, you can set normal, tangent, and binormal to NONE and texcoord0 to half precision (16_16_FLOAT), resulting in only 16 bytes per vertex. Cutting the mesh data in half enables you to load larger models and potentially improves performance.

Memory optimizations

Memory consumption of loaded content may become a bottleneck on the rendering system. If the memory payload becomes too large, it may compromise rendering performance or cause the model to not load altogether. This paragraph discusses some important strategies to reduce the memory footprint.


Instancing is a concept where meshes are reused for parts with distinct spatial transformations, as opposed to every part referencing its own unique geometry. Instancing has significant impact on the memory footprint. Example use cases for instancing are the screws in an engine model or chairs in an architectural model.


Instancing can improve the memory consumption (and thus loading times) significantly, however the improvements on the rendering performance side are insignificant.

The conversion service respects instancing if parts are marked up accordingly in the source file. However, conversion does not perform additional deep analysis of mesh data to identify reusable parts. Thus the content creation tool and its export pipeline are the decisive criteria for proper instancing setup.

A simple way to test whether instancing information gets preserved during conversion is to have a look at the output statistics, specifically the numMeshPartsInstanced member. If the value for numMeshPartsInstanced is larger than zero, it indicates that meshes are shared across instances.

Example: Instancing setup in 3ds Max

Autodesk 3ds Max has distinct object cloning modes called Copy, Instance, and Reference that behave differently with regards to instancing in the exported .fbx file.

Cloning in 3ds Max

  • Copy : In this mode the mesh is cloned, so no instancing is used (numMeshPartsInstanced = 0).
  • Instance : The two objects share the same mesh, so instancing is used (numMeshPartsInstanced = 1).
  • Reference : Distinct modifiers can be applied to the geometries, so the exporter chooses a conservative approach and does not use instancing (numMeshPartsInstanced = 0).

Depth-based composition mode

If memory is a concern, configure the renderer with the depth-based composition mode. In this mode, GPU payload is distributed across multiple GPUs.

Decrease vertex size

As discussed in the best practices for component format changes section, adjusting the vertex format can decrease the memory footprint. However, this option should be the last resort.

Texture sizes

Depending on the type of scenario, the amount of texture data may outweigh the memory used for mesh data. Photogrammetry models are candidates. The conversion configuration does not provide a way to automatically scale down textures. If necessary, texture scaling has to be done as a client-side pre-processing step. The conversion step however does pick a suitable texture compression format:

  • BC1 for opaque color textures
  • BC7 for source color textures with alpha channel

Since format BC7 has twice the memory footprint compared to BC1, it is important to make sure that the input textures do not provide an alpha channel unnecessarily.

Typical use cases

Finding good import settings for a given use case can be a tedious process. On the other hand, conversion settings may have a significant impact on runtime performance.

There are certain classes of use cases that qualify for specific optimizations. Some examples are given below.

Use case: Architectural visualization / large outdoor maps

  • These types of scenes tend to be static, meaning they don't need movable parts. Accordingly, the sceneGraphMode can be set to static or even none, which improves runtime performance. With static mode, the scene's root node can still be moved, rotated, and scaled, for example to dynamically switch between 1:1 scale (for first person view) and a table top view.

  • When you need to move parts around, that typically also means that you need support for raycasts or other spatial queries, so that you can pick those parts in the first place. On the other hand, if you don't intend to move something around, chances are high that you also don't need it to participate in spatial queries and therefore can turn off the generateCollisionMesh flag. This switch has significant impact on conversion times, loading times, and also runtime per-frame update costs.

  • If the application does not use cut planes, the opaqueMaterialDefaultSidedness flag should be turned off. The performance gain is typically 20%-30%. Cut planes can still be used, but there won't be back-faces when looking into the inner parts of objects, which looks counter-intuitive. For more information, see single sided rendering.

Use case: Photogrammetry models

When rendering photogrammetry models there is typically no need for a scene graph, so you could set the sceneGraphMode to none. Since those models rarely contain a complex scene graph to begin with, the impact of this option should be insignificant, though.

Because lighting is already baked into the textures, no dynamic lighting is needed. Therefore:

  • Set the unlitMaterials flag to true to turn all materials into unlit color materials.
  • Remove unneeded data from the vertex format. See the example above.

Use case: Visualization of compact machines, etc.

In these use cases, the models often have very high detail within a small volume. The renderer is heavily optimized to handle such cases well. However, most of the optimizations mentioned in the previous use case do not apply here:

  • Individual parts should be selectable and movable, so the sceneGraphMode must be left to dynamic.
  • Ray casts are typically an integral part of the application, so collision meshes must be generated.
  • Cut planes look better with the opaqueMaterialDefaultSidedness flag enabled.

Deprecated features

Providing settings using the non-model-specific filename conversionSettings.json is still supported but deprecated. Please use the model-specific filename <modelName>.ConversionSettings.json instead.

The use of a material-override setting to identify a Material Override file in the conversion settings file is still supported but deprecated. Please use the model-specific filename <modelName>.MaterialOverrides.json instead.

Next steps