Working with the Asset-Importer-Lib

Importing Data

Access by C++ class interface

The Asset-Importer-Lib can be accessed by both a class or flat function interface. The C++ class the interface is the preferred way of interaction: you create an instance of the class Assimp::Importer, maybe adjust some settings of it and then just call


The class will read the files and process its data, handing back the imported data as a pointer to an aiScene to you. You can now extract the data you need from the file. The importer manages all the resources for itself. If the importer is destroyed, all the data that was created/read by it will be destroyed, too. So the easiest way to use the Importer is to create an instance locally, use its results and then simply let it go out of scope.

C++ example:

#include <assimp/Importer.hpp>      // C++ importer interface
#include <assimp/scene.h>           // Output data structure
#include <assimp/postprocess.h>     // Post processing flags

bool DoTheImportThing( const std::string& pFile) {
  // Create an instance of the Importer class
  Assimp::Importer importer;

  // And have it read the given file with some example postprocessing
  // Usually - if speed is not the most important aspect for you - you'll
  // probably to request more postprocessing than we do in this example.
  const aiScene* scene = importer.ReadFile( pFile,
    aiProcess_CalcTangentSpace       |
    aiProcess_Triangulate            |
    aiProcess_JoinIdenticalVertices  |

  // If the import failed, report it
  if (nullptr == scene) {
    DoTheErrorLogging( importer.GetErrorString());
    return false;

  // Now we can access the file's contents.
  DoTheSceneProcessing( scene);

  // We're done. Everything will be cleaned up by the importer destructor
  return true;

What exactly is read from the files and how you interpret it is described at the:ref:ai_data. The post-processing steps that the Assimp library can apply to the imported data are listed at #aiPostProcessSteps. See the @ref pp Post processing page for more details.

Note that the aiScene data structure returned is declared ‘const’. Yes, you can get rid of these 5 letters with a simple cast. Yes, you may do that. No, it’s not recommended (and it’s suicide in DLL builds if you try to use new or delete on any of the arrays in the scene).

Access by plain-c function interface

The plain function interface is just as simple but requires you to manually call the clean-up after you’re done with the imported data. To start the import process, call aiImportFile() with the filename in question and the desired postprocessing flags like above. If the call is successful, an aiScene pointer with the imported data is handed back to you. When you’re done with the extraction of the data you’re interested in, call aiReleaseImport() on the imported scene to clean up all resources associated with the import.


#include <assimp/cimport.h>        // Plain-C interface
#include <assimp/scene.h>          // Output data structure
#include <assimp/postprocess.h>    // Post processing flags

bool DoTheImportThing( const char* pFile) {
  // Start the import on the given file with some example postprocessing
  // Usually - if speed is not the most important aspect for you - you'll t
  // probably to request more postprocessing than we do in this example.
  const struct aiScene* scene = aiImportFile( pFile,
    aiProcess_CalcTangentSpace       |
    aiProcess_Triangulate            |
    aiProcess_JoinIdenticalVertices  |

  // If the import failed, report it
  if( NULL == scene) {
    DoTheErrorLogging( aiGetErrorString());
    return false;

  // Now we can access the file's contents
  DoTheSceneProcessing( scene);

  // We're done. Release all resources associated with this import
  aiReleaseImport( scene);
  return true;

Using custom IO logic with the C++ class interface

The Assimp library needs to access files internally. This of course applies to the file you want to read, but also to additional files in the same folder for certain file formats. By default, standard C/C++ IO logic is used to access these files. If your application works in a special environment where custom logic is needed to access the specified files, you have to supply custom implementations of IOStream and IOSystem. A shortened example might look like this:

#include <assimp/IOStream.hpp>
#include <assimp/IOSystem.hpp>

// My own implementation of IOStream
class MyIOStream : public Assimp::IOStream {
  friend class MyIOSystem;

  // Constructor protected for private usage by MyIOSystem

  size_t Read( void* pvBuffer, size_t pSize, size_t pCount) { ... }
  size_t Write( const void* pvBuffer, size_t pSize, size_t pCount) { ... }
  aiReturn Seek( size_t pOffset, aiOrigin pOrigin) { ... }
  size_t Tell() const { ... }
  size_t FileSize() const { ... }
  void Flush () { ... }

// Fisher Price - My First Filesystem
class MyIOSystem : public Assimp::IOSystem {
  MyIOSystem() { ... }
  ~MyIOSystem() { ... }

  // Check whether a specific file exists
  bool Exists( const std::string& pFile) const {

  // Get the path delimiter character we'd like to see
  char GetOsSeparator() const {
    return '/';

  // ... and finally a method to open a custom stream
  IOStream* Open( const std::string& pFile, const std::string& pMode) {
    return new MyIOStream( ... );

  void Close( IOStream* pFile) { delete pFile; }

Now that your IO system is implemented, supply an instance of it to the Importer object by calling


An example:

void DoTheImportThing( const std::string& pFile) {
  Assimp::Importer importer;
  // put my custom IO handling in place
  importer.SetIOHandler( new MyIOSystem());

  // the import process will now use this implementation to access any file
  importer.ReadFile( pFile, SomeFlag | SomeOtherFlag);

Using custom IO logic with the plain-c function interface

The C interface also provides a way to override the file system. Control is not as fine-grained as for C++ although surely enough for almost any purpose. The process is simple:

  • Include cfileio.h

  • Fill an aiFileIO structure with custom file system callbacks (they’re self-explanatory as they work similarly to the CRT’s fXXX functions)

  • and pass it as a parameter to #aiImportFileEx


The Assimp-library provides an easy mechanism to log messages. For instance, if you want to check the state of your import and you just want to see, after which preprocessing step the import process was aborted you can take a look into the log. Per default, the Assimp-library provides a default log implementation, where you can log your user-specific message by calling it a singleton with the requested logging type. To see how this works take a look at this:

using namespace Assimp;

// Create a logger instance
DefaultLogger::create("", Logger::VERBOSE);

// Now I am ready for logging my stuff
DefaultLogger::get()->info("this is my info-call");

// Kill it after the work is done

At first, you have to create the default-logger-instance (create). Now you are ready to rock and can log a little bit around. After that, you should kill it to release the singleton instance.

If you want to integrate the assimp-log into your own GUI it may be helpful to have a mechanism writing the logs into your own log windows. The logger interface provides this by implementing an interface called LogStream. You can attach and detach this log stream to the default-logger instance or any implementation derived from Logger. Just derivate your own logger from the abstract base-class LogStream and overwrite the write method:

// Example stream
class myStream : public LogStream {
    // Write something using your own functionality
    void write(const char* message) {
        ::printf("%s\n", message);

// Select the kinds of messages you want to receive on this log stream
const unsigned int severity = Logger::Debugging|Logger::Info|Logger::Err|Logger::Warn;

// Attaching it to the default logger
Assimp::DefaultLogger::get()->attachStream( new myStream, severity );

The severity level controls the kind of message which will be written into the attached stream. If you just want to log errors and warnings set the warn and error severity flag for those severities. It is also possible to remove a self-defined log stream from an error severity by detaching it with the severity flag set:

unsigned int severity = 0;
severity |= Logger::Debugging;

// Detach debug messages from your self defined stream
Assimp::DefaultLogger::get()->attachStream( new myStream, severity );

If you want to implement your own logger just derive from the abstract base class Logger and overwrite the methods debug, info, warn, and error.

If you want to see the debug messages in a debug-configured build, the Logger-interface provides a logging severity. You can set it by calling the following method:

Assimp::DefaultLogger::get()->setLogSeverity( LogSeverity log_severity );

The normal logging severity supports just the basic stuff like info, warnings, and errors. In the verbose level very fine-grained debug messages will be logged, too. Note that this kind of logging might decrease import performance.

Exporting models

A valid aiScene instance can be used to export into a requested assset-format.

For instance:

bool exporterTest() override {
    ::Assimp::Importer importer;
    ::Assimp::Exporter exporter;
    const aiScene *scene = importer.ReadFile(ASSIMP_TEST_MODELS_DIR "/OBJ/spider.obj", aiProcess_ValidateDataStructure);
    exporter.Export(scene, "obj", ASSIMP_TEST_MODELS_DIR "/OBJ/spider_out.obj");
    return true;

The following file formats are currently supported:

  • collada

  • x

  • stp

  • obj

  • objnomtl

  • stl

  • stlb

  • ply

  • plyb

  • 3ds

  • gltf2

  • glb2

  • gltf

  • glb

  • assbin

  • assxml

  • x3d

  • fbx

  • fbxa

  • m3da

  • 3mf

  • pbrt

  • assjson

Data Structures


The Assimp-Library returns the imported data in a collection of structures. aiScene forms the root of the data, from here you gain access to all the nodes, meshes, materials, animations, or textures that was read from the imported file. The aiScene is returned from a successful call to Assimp::Importer::ReadFile(), aiImportFile() or aiImportFileEx() - see the ai_usage for further information on how to use the library.

By default, all 3D data is provided in a right-handed coordinate system such as OpenGL uses. In this coordinate system, +X points to the right, +Y points upwards and +Z points out of the screen towards the viewer. Several modeling packages such as 3D Studio Max use this coordinate system as well (or a rotated variant of it). By contrast, some other environments use left-handed coordinate systems, a prominent example being DirectX. If you need the imported data to be in a left-handed coordinate system, supply the #aiProcess_MakeLeftHanded flag to the ReadFile() function call.

The output face winding is counterclockwise. Use #aiProcess_FlipWindingOrder to get CW data.



Outputted polygons can be literally everything: they’re probably concave, self-intersecting or non-planar, although our built-in triangulation (#aiProcess_Triangulate postprocessing step) doesn’t handle the two latter.

The output UV coordinate system has its origin in the lower-left corner:

0x|1y ---------- 1x|1y
 |                |
 |                |
 |                |
0x|0y ---------- 1x|0y

Use the #aiProcess_FlipUVs flag to get UV coordinates with the upper-left corner als origin.

A typical 4x4 matrix including a translational part looks like this:

X1  Y1  Z1  T1
X2  Y2  Z2  T2
X3  Y3  Z3  T3
 0   0   0   1

with <tt>(X1, X2, X3)</tt> being the local X base vector, <tt>(Y1, Y2, Y3)</tt> being the local Y base vector, <tt>(Z1, Z2, Z3)</tt> being the local Z base vector and <tt>(T1, T2, T3)</tt> being the offset of the local origin (the translational part). All matrices in the library use row-major storage order. That means that the matrix elements are stored row-by-row, i.e. they end up like this in memory: <tt>[X1, Y1, Z1, T1, X2, Y2, Z2, T2, X3, Y3, Z3, T3, 0, 0, 0, 1]</tt>.

Note that this is neither the OpenGL format nor the DirectX format, because both of them specify the matrix layout such that the translational part occupies three consecutive addresses in memory (so those matrices end with <tt>[…, T1, T2, T3, 1]</tt>), whereas the translation in an Assimp matrix is found at the offsets 3, 7 and 11 (spread across the matrix). You can transpose an Assimp matrix to end up with the format that OpenGL and DirectX mandate. To be very precise: The transposition has nothing to do with a left-handed or right-handed coordinate system but ‘converts’ between row-major and column-major storage formats.

<b>11.24.09:</b> We changed the orientation of our quaternions to the most common convention to avoid confusion. However, if you’re a previous user of Assimp and you update the library to revisions beyond SVNREV 502, you have to adapt your animation loading code to match the new quaternion orientation.

The Node-Hierarchy

Nodes are little-named entities in the scene that have a place and orientation relative to their parents. Starting from the scene’s root node all nodes can have 0 to x child nodes, thus forming a hierarchy. They form the base on which the scene is built: a node can refer to 0..x meshes, can be referred to by a bone of a mesh, or can be animated by a key sequence of animation. DirectX calls them “frames”, others call them “objects”, and we call them aiNode.

A node can potentially refer to single or multiple meshes. The meshes are not stored inside the node, but instead in an array of aiMesh inside the aiScene. A node only refers to them by their array index. This also means that multiple nodes can refer to the same mesh, which provides a simple form of instancing. A mesh referred to in this way lives in the node’s local coordinate system. If you want the mesh’s orientation in global space, you’d have to concatenate the transformations from the referring node and all of its parents.

Most of the file formats don’t really support complex scenes, though, but a single model only. But there are more complex formats such as .3ds, .x, or .collada scenes which may contain an arbitrarily complex hierarchy of nodes and meshes. I myself would suggest a recursive filter function such as the following pseudocode:

void CopyNodesWithMeshes( aiNode node, SceneObject targetParent, Matrix4x4 accTransform) {
  SceneObject parent;
  Matrix4x4 transform;

  //If node has meshes, create a new scene object for it
  if( node.mNumMeshes > 0) {
        SceneObjekt newObject = new SceneObject;
        targetParent.addChild( newObject);
        // copy the meshes
        CopyMeshes( node, newObject);

        //The new object is the parent for all child nodes
        parent = newObject;
  } else {
        // if no meshes, skip the node, but keep its transformation
        parent = targetParent;
        transform = node.mTransformation * accTransform;

  // continue for all child nodes
  for( all node.mChildren) {
        CopyNodesWithMeshes( node.mChildren[a], parent, transform);

This function copies a node into the scene graph if it has children. If yes, a new scene object is created for the import node and the node’s meshes are copied over. If not, no object is created. Potential child objects will be added to the old targetParent, but their transformation will be correct in respect to the global space. This function also works great in filtering the bone nodes - nodes that form the bone hierarchy for another mesh/node, but don’t have any mesh themselves.


All meshes of an imported scene are stored in an array of aiMesh* inside the aiScene. Nodes refer to them by their index in the array and provide the coordinate system for them, too. One mesh uses only a single material everywhere - if parts of the model use a different material, this part is moved to a separate mesh at the same node. The mesh refers to its material in the same way as the node refers to its meshes: materials are stored in an array inside aiScene, and the mesh stores only an index into this array.

An aiMesh is defined by a series of data channels. The presence of these data channels is defined by the contents of the imported file: by default, there are only those data channels present in the mesh that were also found in the file. The only channels guaranteed to be always present are aiMesh::mVertices and aiMesh::mFaces. You can test for the presence of other data by testing the pointers against NULL or using the helper functions provided by aiMesh. You may also specify several post-processing flags at Importer::ReadFile() to let Assimp calculate or recalculate additional data channels for you.

At the moment, a single aiMesh may contain a set of triangles and polygons. A single vertex does always have a position. In addition, it may have one normal, one tangent, and bitangent, zero to AI_MAX_NUMBER_OF_TEXTURECOORDS (4 at the moment) texture coords and zero to AI_MAX_NUMBER_OF_COLOR_SETS (4) vertex colors. In addition, a mesh may or may not have a set of bones described by an array of aiBone structures. How to interpret the bone information is described later on.


See the:ref:ai_material Material System Page.


A mesh may have a set of bones in the form of instances from the aiBone objects. Bones are a means to deform a mesh according to the movement of a skeleton. Each bone has a name and a set of vertices on which it has influence. Its offset matrix declares the transformation needed to transform from mesh space to the local space of this bone.

Using the bones name you can find the corresponding node in the node hierarchy. This node in relation to the other bones’ nodes defines the skeleton of the mesh. Unfortunately, there might also be nodes that are not used by a bone in the mesh but still affect the pose of the skeleton because they have child nodes which are bones. So when creating the skeleton hierarchy for a mesh I suggest the following method:

a. Create a map or a similar container to store which nodes are necessary for the skeleton. Pre-initialize it for all nodes with a “no”.

  1. For each bone in the mesh:

    b1. Find the corresponding node in the scene’s hierarchy by comparing their names.

    b2. Mark this node as “yes” in the necessityMap.

    b3. Mark all of its parents the same way until you

    1. find the mesh’s node or

    2. the parent of the mesh’s node.

  2. Recursively iterate over the node hierarchy.

c1. If the node is marked as necessary, copy it into the skeleton and check its children.

c2. If the node is marked as not necessary, skip it and do not iterate over its children.

Reasons: you need all the parent nodes to keep the transformation chain intact. For most file formats and modeling packages, the node hierarchy of the skeleton is either a child of the mesh node or a sibling of the mesh node but this is by no means a requirement so you shouldn’t rely on it. The node closest to the root node is your skeleton root, from there, you start copying the hierarchy. You can skip every branch without a node being a bone in the mesh - that’s why the algorithm skips the whole branch if the node is marked as “not necessary”.

You should now have a mesh in your engine with a skeleton that is a subset of the imported hierarchy.


An imported scene may contain zero to n aiAnimation entries. An animation in this context is a set of keyframe sequences where each sequence describes the orientation of a single node in the hierarchy over a limited time span. Animations of this kind are usually used to animate the skeleton of a skinned mesh, but there are other uses as well.

An aiAnimation has a duration. The duration as well as all time stamps are given in ticks. To get the correct timing, all time stamps thus have to be divided by aiAnimation::mTicksPerSecond. Beware, though, that certain combinations of file format and exporter don’t always store this information in the exported file. In this case, mTicksPerSecond is set to 0 to indicate the lack of knowledge.

The aiAnimation consists of a series of aiNodeAnim. Each bone animation affects a single node in the node hierarchy only, the name specifying which node is affected. For this node the structure stores three separate key sequences: a vector key sequence for the position, a quaternion key sequence for the rotation, and another vector key sequence for the scaling. All 3d data is local to the coordinate space of the node’s parent, which means in the same space as the node’s transformation matrix. There might be cases where animation tracks refer to a non-existent node by their name, but this should not be the case in your everyday data.

To apply such an animation you need to identify the animation tracks that refer to actual bones in your mesh. Then for every track: * Find the keys that lay right before the current anim time. * Optional: interpolate between these and the following keys. * Combine the calculated position, rotation, and scaling into a transformation matrix * Set the affected node’s transformation to the calculated matrix.

If you need hints on how to convert to or from quaternions, have a look at the Matrix & Quaternion FAQ. I suggest using logarithmic interpolation for the scaling keys if you happen to need them - usually, you don’t need them at all.




Normally textures used by assets are stored in separate files, however, there are file formats embedding their textures directly into the model file. Such textures are loaded into an aiTexture structure.

In previous versions, the path from the query for AI_MATKEY_TEXTURE(textureType, index) would be *<index> where <index> is the index of the texture in aiScene::mTextures. Now this call will return a file path for embedded textures in FBX files. To test if it is an embedded texture use aiScene::GetEmbeddedTexture. If the returned pointer is not null, it is embedded and can be loaded from the data structure. If it is null, search for a separate file. Other file types still use the old behavior.

If you rely on the old behavior, you can use Assimp::Importer::SetPropertyBool with the key #AI_CONFIG_IMPORT_FBX_EMBEDDED_TEXTURES_LEGACY_NAMING to force the old behavior.

There are two cases:
  • The texture is NOT compressed. Its color data is directly stored in the aiTexture structure as an array of aiTexture::mWidth * aiTexture::mHeight aiTexel structures. Each aiTexel represents a pixel (or “texel”) of the texture image. The color data is stored in an unsigned RGBA8888 format, which can be easily used for both Direct3D and OpenGL (swizzling the order of the color components might be necessary). RGBA8888 has been chosen because it is well-known, easy to use , and natively supported by nearly all graphics APIs.

  • This applies if aiTexture::mHeight == 0 is fulfilled. Then, the texture is stored in a compressed format such as DDS or PNG. The term “compressed” does not mean that the texture data must actually be compressed, however, the texture was found in the model file as if it was stored in a separate file on the hard disk. Appropriate decoders (such as libjpeg, libpng, D3DX, DevIL) are required to load these textures. aiTexture::mWidth specifies the size of the texture data in bytes, aiTexture::pcData is a pointer to the raw image data and aiTexture::achFormatHint is either zeroed or contains the most common file extension of the embedded texture’s format. This value is only set if Assimp is able to determine the file format.


All materials are stored in an array of aiMaterial inside the aiScene.

Each aiMesh refers to one material by its index in the array. Due to the vastly diverging definitions and usages of material parameters, there is no hard definition of a material structure. Instead, a material is defined by a set of properties accessible by their names. Have a look at assimp/material.h to see what types of properties are defined. In this file, there are also various functions defined to test for the presence of certain properties in a material and retrieve their values.


Textures are organized in stacks, each stack being evaluated independently. The final color value from a particular texture stack is used in the shading equation. For example, the computed color value of the diffuse texture stack (aiTextureType_DIFFUSE) is multiplied with the amount of incoming diffuse light to obtain the final diffuse color of a pixel.


Resulting equation

Constant base color


Blend operation 0

Strength factor 0


Texture 0


Blend operation 1


Strength factor 1


Texture 1



All material key constants start with ‘AI_MATKEY’ as a prefix.


Data Type

Default Value






The name of the material, if available.

Ignored by <tt>aiProcess_RemoveRedundantMaterials. Materials are considered equal even if their names are different.



black (0,0,0)

Diffuse color of the material. This is typically scaled by the amount of incoming diffuse light (e.g. using gouraud shading)




black (0,0,0)

Specular color of the material. This is typically scaled by the amount of incoming specular light (e.g. using phong shading)




black (0,0,0)

Ambient color of the material. This is typically scaled by the amount of ambient light




black (0,0,0)

Emissive color of the material. This is the amount of light emitted by the object. In real time applications it will usually not affect surrounding objects, but raytracing applications may wish to treat emissive objects as light sources.




black (0,0,0)

Defines the transparent color of the material, this is the color to be multiplied with the color of translucent light to construct the final ‘destination color’ for a particular position in the screen buffer.




<td>black (0,0,0)

Defines the reflective color of the material. This is typically scaled by the amount of incoming light from the direction of mirror reflection. Usually combined with an environment lightmap of some kind for real-time applications.





Scales the reflective color of the material.





Specifies whether wireframe rendering must be turned on for the material. 0 for false, !0 for true.





Specifies whether meshes using this material must be rendered without backface culling. 0 for false, !0 for true.

Some importers set this property if they don’t know whether the output face order is right. As long as it is not set, you may safely enable backface culling.




One of the #aiShadingMode enumerated values. Defines the library shading model to use for (real time) rendering to approximate the original look of the material as closely as possible.

The presence of this key might indicate a more complex material. If absent, assume phong shading only if a specular exponent is given.




One of the #aiBlendMode enumerated values. Defines how the final color value in the screen buffer is computed from the given color at that position and the newly computed color from the material. Simply said, alpha blending settings.





Defines the opacity of the material in a range between 0..1.

Use this value to decide whether you have to activate alpha blending for rendering. OPACITY!=1 usually also implies TWOSIDED=1 to avoid cull artifacts.




Defines the shininess of a phong-shaded material. This is actually the exponent of the phong specular equation

SHININESS=0 is equivalent to SHADING_MODEL = aiShadingMode_Gouraud




Scales the specular color of the material.

This value is kept separate from the specular color by most modelers, and so do we.




Defines the Index Of Refraction for the material. That’s not supported by most file formats.

Might be of interest for raytracing.




Defines the path of the n’th texture on the stack ‘t’, where ‘n’ is any value >= 0 and ‘t’ is one of the #aiTextureType enumerated values. A file path to an external file or an embedded texture. Use aiScene::GetEmbeddedTexture to test if it is embedded for FBX files, in other cases embedded textures start with ‘*’ followed by an index into aiScene::mTextures.

See the @ref mat_tex section above. Also see @ref textures for a more information about texture retrieval.




Defines the strength the n’th texture on the stack ‘t’. All color components (rgb) are multiplied with this factor before any further processing is done.





One of the #aiTextureOp enumerated values. Defines the arithmetic operation to be used to combine the n’th texture on the stack ‘t’ with the n-1’th. TEXOP(t,0) refers to the blend operation between the base color for this stack (e.g. COLOR_DIFFUSE for the diffuse stack) and the first texture.





Defines how the input mapping coordinates for sampling the n’th texture on the stack ‘t’ are computed. Usually explicit UV coordinates are provided, but some model file formats might also be using basic shapes, such as spheres or cylinders, to project textures onto meshes.

See the ‘Textures’ section below. #aiProcess_GenUVCoords can be used to let Assimp compute proper UV coordinates from projective mappings.




Defines the UV channel to be used as input mapping coordinates for sampling the n’th texture on the stack ‘t’. All meshes assigned to this material share the same UV channel setup

Presence of this key implies MAPPING(t,n) to be #aiTextureMapping_UV. See @ref uvwsrc for more details.




Any of the #aiTextureMapMode enumerated values. Defines the texture wrapping mode on the x axis for sampling the n’th texture on the stack ‘t’. ‘Wrapping’ occurs whenever UVs lie outside the 0..1 range.





Wrap mode on the v axis. See MAPPINGMODE_U.





Defines the base axis to to compute the mapping coordinates for the n’th texture on the stack ‘t’ from. This is not required for UV-mapped textures. For instance, if <tt>MAPPING(t,n)</tt> is #aiTextureMapping_SPHERE, U and V would map to longitude and latitude of a sphere around the given axis. The axis is given in local mesh space.





Defines miscellaneous flag for the n’th texture on the stack ‘t’. This is a bitwise combination of the #aiTextureFlags enumerated values.



Retrieving a property from a material is done using various utility functions. For C++ it’s simply calling aiMaterial::Get()

aiMaterial* mat = .....

// The generic way
if(AI_SUCCESS != mat->Get(<material-key>,<where-to-store>)) {
   // handle epic failure here

Simple, isn’t it? To get the name of a material you would use

aiString name;

Or for the diffuse color (‘color’ won’t be modified if the property is not set)

aiColor3D color (0.f,0.f,0.f);

<b>Note:</b> Get() is actually a template with explicit specializations for aiColor3D, aiColor4D, aiString, float, int and some others. Make sure that the type of the second parameter matches the expected data type of the material property (no compile-time check yet!). Don’t follow this advice if you wish to encounter very strange results.


For good old C it’s slightly different. Take a look at the aiGetMaterialGet<data-type> functions.

aiMaterial* mat = .....

if(AI_SUCCESS != aiGetMaterialFloat(mat,<material-key>,<where-to-store>)) {
   // handle epic failure here

To get the name of a material you would use

aiString name;

Or for the diffuse color (‘color’ won’t be modified if the property is not set)

aiColor3D color (0.f,0.f,0.f);

How to map UV channels to textures (MATKEY_UVWSRC)

The MATKEY_UVWSRC property is only present if the source format doesn’t specify an explicit mapping from textures to UV channels. Many formats don’t do this and assimp is not aware of a perfect rule either.

Your handling of UV channels needs to be flexible therefore. Our recommendation is to use logic like this to handle most cases properly:

have only one uv channel?

assign channel 0 to all textures and break

for all textures
have uvwsrc for this texture?

assign channel specified in uvwsrc


assign channels in ascending order for all texture stacks, i.e. diffuse1 gets channel 1, opacity0 gets channel 0.

Pseudo Code Listing

For completeness, the following is a very rough pseudo-code sample showing how to evaluate Assimp materials in your shading pipeline. You’ll probably want to limit your handling of all those material keys to a reasonable subset suitable for your purposes (for example most 3d engines won’t support highly complex multi-layer materials, but many 3d modellers do).

Also note that this sample is targeted at a (shader-based) rendering pipeline for real-time graphics.

// ---------------------------------------------------------------------------------------
// Evaluate multiple textures stacked on top of each other
float3 EvaluateStack(stack) {
  // For the 'diffuse' stack.base_color would be COLOR_DIFFUSE
  // and TEXTURE(aiTextureType_DIFFUSE,n) the n'th texture.

  float3 base = stack.base_color;
  for (every texture in the stack) {
    // assuming we have explicit & pre transformed UVs for this texture
    float3 color = SampleTexture(texture,uv);

    // scale by texture blend factor
    color *= texture.blend;

    if (texture.op == add)
      base += color;
    else if (texture.op == multiply)
      base *= color;
    else // other blend ops go here
  return base;

// ---------------------------------------------------------------------------------------
// Compute the diffuse contribution for a pixel
float3 ComputeDiffuseContribution() {
  if (shading == none)
     return float3(1,1,1);

  float3 intensity (0,0,0);
  for (all lights in range) {
    float fac = 1.f;
    if (shading == gouraud)
      fac =  lambert-term ..
    else // other shading modes go here

    // handling of different types of lights, such as point or spot lights
    // ...

    // and finally sum the contribution of this single light ...
    intensity += light.diffuse_color * fac;
  // ... and combine the final incoming light with the diffuse color
  return EvaluateStack(diffuse) * intensity;

// ---------------------------------------------------------------------------------------
// Compute the specular contribution for a pixel
float3 ComputeSpecularContribution() {
  if (shading == gouraud || specular_strength == 0 || specular_exponent == 0)
    return float3(0,0,0);

  float3 intensity (0,0,0);
  for (all lights in range) {
    float fac = 1.f;
    if (shading == phong)
      fac =  phong-term ..
    else // other specular shading modes go here

    // handling of different types of lights, such as point or spot lights
    // ...

    // and finally sum the specular contribution of this single light ...
    intensity += light.specular_color * fac;
  // ... and combine the final specular light with the specular color
  return EvaluateStack(specular) * intensity * specular_strength;

// ---------------------------------------------------------------------------------------
// Compute the ambient contribution for a pixel
float3 ComputeAmbientContribution() {
  if (shading == none)
     return float3(0,0,0);

  float3 intensity (0,0,0);
  for (all lights in range) {
    float fac = 1.f;

    // handling of different types of lights, such as point or spot lights
    // ...

    // and finally sum the ambient contribution of this single light ...
    intensity += light.ambient_color * fac;
  // ... and combine the final ambient light with the ambient color
  return EvaluateStack(ambient) * intensity;

// ---------------------------------------------------------------------------------------
// Compute the final color value for a pixel
// @param prev Previous color at that position in the framebuffer
float4 PimpMyPixel (float4 prev) {
  // .. handle displacement mapping per vertex
  // .. handle bump/normal mapping

  // Get all single light contribution terms
  float3 diff = ComputeDiffuseContribution();
  float3 spec = ComputeSpecularContribution();
  float3 ambi = ComputeAmbientContribution();

  // .. and compute the final color value for this pixel
  float3 color = diff + spec + ambi;
  float3 opac  = EvaluateStack(opacity);

  // note the *slightly* strange meaning of additive and multiplicative blending here ...
  // those names will most likely be changed in future versions
  if (blend_func == add)
       return prev+color*opac;
  else if (blend_func == multiply)
       return prev*(1.0-opac)+prev*opac;

   return color;

How to access shader-code

You can get the shadercode like a texture (AI_MATKEY_GLOBAL_SHADERLANG and AI_MATKEY_SHADER_VERTEX, …) You can get assigned shader sources by using the following material keys:

  • AI_MATKEY_GLOBAL_SHADERLANG To get the used shader language.

  • AI_MATKEY_SHADER_VERTEX Assigned vertex shader code stored as a string.

  • AI_MATKEY_SHADER_FRAGMENT Assigned fragment shader code stored as a string.

  • AI_MATKEY_SHADER_GEO Assigned geometry shader code stored as a string.

  • AI_MATKEY_SHADER_TESSELATION Assigned tesselation shader code stored as a string.

  • AI_MATKEY_SHADER_PRIMITIVE Assigned primitive shader code stored as a string.

  • AI_MATKEY_SHADER_COMPUTE Assigned compute shader code stored as a string.

Animation Overview

<a href=””>This external tutorial</a> has working code to get started implementing animations using bone matrix array in the vertex shader. (If using glm (OpenGL math library), cross-reference with <a href=””>this page</a> which has useful tips on converting between Assimp and glm objects)


This diagram shows how you can calculate your transformation matrices for an animated character: .. image:: ../images/AnimationOverview.png



This page discusses general performance issues related to Assimp.


Assimp has built-in support for <i>very</i> basic profiling and time measurement. To turn it on, set the <tt>GLOB_MEASURE_TIME</tt> configuration switch to <tt>true</tt> (nonzero). Results are dumped to the log file, so you need to set up an appropriate logger implementation with at least one output stream first (see the @:ref:ai_logging for the details.).

Note that these measurements are based on a single run of the importer and each of the post-processing steps, so a single result set is far away from being significant in a statistical sense. While precision can be improved by running the test multiple times, the low accuracy of the timings may render the results useless for smaller files.

A sample report looks like this (some unrelated log messages omitted, entries grouped for clarity):

Debug, T5488: START `total`
Info,  T5488: Found a matching importer for this file format

Debug, T5488: START `import`
Info,  T5488: BlendModifier: Applied the `Subdivision` modifier to `OBMonkey`
Debug, T5488: END   `import`, dt= 3.516 s

Debug, T5488: START `preprocess`
Debug, T5488: END   `preprocess`, dt= 0.001 s
Info,  T5488: Entering post processing pipeline

Debug, T5488: START `postprocess`
Debug, T5488: RemoveRedundantMatsProcess begin
Debug, T5488: RemoveRedundantMatsProcess finished
Debug, T5488: END   `postprocess`, dt= 0.001 s

Debug, T5488: START `postprocess`
Debug, T5488: TriangulateProcess begin
Info,  T5488: TriangulateProcess finished. All polygons have been triangulated.
Debug, T5488: END   `postprocess`, dt= 3.415 s

Debug, T5488: START `postprocess`
Debug, T5488: SortByPTypeProcess begin
Info,  T5488: Points: 0, Lines: 0, Triangles: 1, Polygons: 0 (Meshes, X = removed)
Debug, T5488: SortByPTypeProcess finished

Debug, T5488: START `postprocess`
Debug, T5488: JoinVerticesProcess begin
Debug, T5488: Mesh 0 (unnamed) | Verts in: 503808 out: 126345 | ~74.922
Info,  T5488: JoinVerticesProcess finished | Verts in: 503808 out: 126345 | ~74.9
Debug, T5488: END   `postprocess`, dt= 2.052 s

Debug, T5488: START `postprocess`
Debug, T5488: FlipWindingOrderProcess begin
Debug, T5488: FlipWindingOrderProcess finished
Debug, T5488: END   `postprocess`, dt= 0.006 s

Debug, T5488: START `postprocess`
Debug, T5488: LimitBoneWeightsProcess begin
Debug, T5488: LimitBoneWeightsProcess end
Debug, T5488: END   `postprocess`, dt= 0.001 s

Debug, T5488: START `postprocess`
Debug, T5488: ImproveCacheLocalityProcess begin
Debug, T5488: Mesh 0 | ACMR in: 0.851622 out: 0.718139 | ~15.7
Info,  T5488: Cache relevant are 1 meshes (251904 faces). Average output ACMR is 0.718139
Debug, T5488: ImproveCacheLocalityProcess finished.
Debug, T5488: END   `postprocess`, dt= 1.903 s

Info,  T5488: Leaving post processing pipeline
Debug, T5488: END   `total`, dt= 11.269 s

In this particular example only one fourth of the total import time was spent on the actual importing, while the rest of the time got consumed by the #aiProcess_Triangulate, #aiProcess_JoinIdenticalVertices and #aiProcess_ImproveCacheLocality postprocessing steps. A wise selection of postprocessing steps is therefore essential to getting good performance. Of course this depends on the individual requirements of your application, in many of the typical use cases of assimp performance won’t matter (i.e. in an offline content pipeline).


You can use the Asset-Importer-Library in a separate thread context. Just make sure that the resources used by the thread are not shared. At this moment, assimp will not make sure that it is safe over different thread contexts.


This page discusses both Assimp scalability in threaded environments and the precautions to be taken in order to use it from multiple threads concurrently.

Thread-safety / using Assimp concurrently from several threads

The library can be accessed by multiple threads simultaneously, as long as the following prerequisites are fulfilled:

  • Users of the C++-API should ensure that they use a dedicated #Assimp::Importer instance for each thread. Constructing instances of #Assimp::Importer is expensive, so it might be a good idea to let every thread maintain its own thread-local instance (which can be used to load as many files as necessary).

  • The C-API is thread-safe.

  • When supplying custom IO logic, one must make sure the underlying implementation is thread-safe.

  • Custom log streams or logger replacements have to be thread-safe, too.

Multiple concurrent imports may or may not be beneficial, however. For certain file formats in conjunction with little or no post-processing IO times tend to be the performance bottleneck. Intense post-processing together with ‘slow’ file formats like X or Collada might scale well with multiple concurrent imports.

Internal threading

Internal multi-threading is not currently implemented.


This page lists some useful resources for Assimp. Note that, even though the core team has an eye on them, we cannot guarantee the accuracy of third-party information. If in doubt, it’s best to ask either on the mailing list or on our forums on

  • Assimp comes with some sample applications, these can be found in the <i>./samples</i> folder. Don’t forget to read the <i>README</i> file.

  • Assimp-GL-Demo - OpenGl animation sample using the library’s animation import facilities.

  • Assimp-Animation-Loader is another utility to simplify animation playback.

  • Assimp-Animations - Tutorial “Loading models using the Open Asset Import Library”, out of a series of OpenGL tutorials.

Importer Notes


This section contains implementation notes for the Blender3D importer.


Assimp provides a self-contained reimplementation of Blender’s so-called SDNA system ( `Notes on SDNA`_ ). SDNA allows Blender to be fully backward and forward-compatible and to exchange files across all platforms. The BLEND format is thus a non-trivial binary monster and the loader tries to read the most of it, naturally limited by the scope of the #aiScene output data structure. Consequently, if Blender is the only modeling tool in your asset workflow, consider writing a custom exporter from Blender if Assimp format coverage does not meet the requirements.

Current status

The Blender loader does not support animations yet, but is apart from that considered relatively stable.

@subsection bl_notes Notes

When filing bugs on the Blender loader, always give the Blender version (or, even better, post the file that caused the error).


This section contains implementation notes on the IFC-STEP importer.


The library provides a partial implementation of the IFC2x3 industry standard for automatized exchange of CAE/architectural data sets. See IFC for more information on the format. We aim at getting as much 3D data out of the files as possible.

Current status

IFC support is new and considered experimental. Please report any bugs you may encounter.


  • Only the STEP-based encoding is supported. IFCZIP and IFCXML are not (but IFCZIP can simply be unzipped to get a STEP file).

  • The importer leaves vertex coordinates untouched, but applies a global scaling to the root transform to convert from whichever unit the IFC file uses to <i>metres</i>.

  • If multiple geometric representations are provided, the choice of which one to load is based on how expensive a representation seems to be in terms of import time. The loader also avoids representation types for which it has known deficits.

  • Not supported are arbitrary binary operations (binary clipping is implemented, though).

  • Of the various relationship types that IFC knows, only aggregation, containment and material assignment are resolved and mapped to the output graph.

  • The implementation knows only about IFC2X3 and applies this rule set to all models it encounters, regardless of their actual version. Loading of older or newer files may fail with parsing errors.


IFC file properties (IfcPropertySet) are kept as per-node metadata, see aiNode::mMetaData.


ATTENTION: The Ogre-Loader is currently under development, many things have changed after this documentation was written, but they are not final enough to rewrite the documentation. So things may have changed by now!

This section contains implementations notes for the OgreXML importer.


Ogre importer is currently optimized for the Blender Ogre exporter because that’s the only one that I use. You can find the Blender Ogre exporter at: OGRE3D forum

What will be loaded?

Mesh: Faces, Positions, Normals, and all TexCoords. The Materialname will be used to load the material.

Material: The right material in the file will be searched, the importer should work with materials who have 1 technique and 1 pass in this technique. From there, the texture name (for 1 color- and 1 normal map) and the materialcolors (but not in custom materials) will be loaded. Also, the materialname will be set.

Skeleton: Skeleton with Bone hierarchy (Position and Rotation, but no Scaling in the skeleton is supported), names and transformations, animations with rotation, translation and scaling keys.

How to export Files from Blender

You can find information about how to use the Ogreexporter on your own, so here are just some options that you need, so the assimp importer will load everything correctly:

  • Use either “Rendering Material” or “Custom Material” see @ref material

  • do not use “Flip Up Axies to Y”

  • use “Skeleton name follow mesh”


There is a binary and an XML mesh Format from Ogre. This loader can only Handle XML files, but don’t panic, there is a command line converter, which you can use to create XML files from Binary Files. Just look on the Ogre page for it.

Currently, you can only load meshes. So you will need to import the .mesh.xml file, the loader will try to find the appendant material and skeleton file.

The skeleton file must have the same name as the mesh file, e.g. fish.mesh.xml and fish.skeleton.xml.

@subsection material Materials The material file can have the same name as the mesh file (if the file is model.mesh or model.mesh.xml the loader will try to load model.material), or you can use

Importer::Importer::SetPropertyString(AI_CONFIG_IMPORT_OGRE_MATERIAL_FILE, "materiafile.material")

to specify the name of the material file. This is especially useful if multiple materials a stored in a single file. The importer will first try to load the material with the same name as the mesh and only if this can’t be opened try to load the alternate material file. The default material filename is “Scene.material”.

We suggest that you use custom materials, because they support multiple textures (like colormap and normalmap). First of all you should read the custom material sektion in the Ogre Blender exporter Help File, and than use the assimp.tlp template, which you can find in scripts/OgreImpoter/Assimp.tlp in the assimp source. If you don’t set all values, don’t worry, they will be ignored during import.

If you want more properties in custom materials, you can easily expand the ogre material loader, it will be just a few lines for each property. Just look in OgreImporterMaterial.cpp


  • IMPORT_OGRE_TEXTURETYPE_FROM_FILENAME: Normally, a texture is loaded as a colormap, if no target is specified in the materialfile. Is this switch is enabled, texture names ending with _n, _l, _s are used as normalmaps, lightmaps or specularmaps.

    Property type: Bool. Default value: false.

  • IMPORT_OGRE_MATERIAL_FILE: Ogre Meshes contain only the MaterialName, not the MaterialFile. If there is no material file with the same name as the material, Ogre Importer will try to load this file and search the material in it.

    Property type: String. Default value: guessed.


You can use properties to chance the behavior of you importer. In order to do so, you have to override BaseImporter::SetupProperties, and specify you custom properties in config.h. Just have a look to the other AI_CONFIG_IMPORT_* defines and you will understand, how it works.

The properties can be set with Importer::SetProperty***() and can be accessed in your SetupProperties function with Importer::GetProperty***(). You can store the properties as a member variable of your importer, they are thread safe.

Notes for text importers

  • Try to make your parser as flexible as possible. Don’t rely on particular layout, whitespace/tab style, except if the file format has a strict definition, in which case you should always warn about spec violations. But the general rule of thumb is be strict in what you write and tolerant in what you accept.

  • Call Assimp::BaseImporter::ConvertToUTF8() before you parse anything to convert foreign encodings to UTF-8. That’s not necessary for XML importers, which must use the provided XML-Parser for reading.

Notes for binary importers

  • Take care of endianness issues! Assimp importers mostly support big-endian platforms, which define the <tt>AI_BUILD_BIG_ENDIAN</tt> constant. See the next section for a list of utilities to simplify this task.

  • Don’t trust the input data! Check all offsets!


Mixed stuff for internal use by loaders, mostly documented (most of them are already included by AssimpPCH.h):

  • ByteSwapper (ByteSwapper.h) - manual byte swapping stuff for binary loaders.

  • StreamReader (StreamReader.h) - safe, endianess-correct, binary reading.

  • XmlParser (XmlParser.hh) - The XML-Parser used in Asset-importer-Lib

  • CommentRemover (RemoveComments.h) - remove single-line and multi-line comments from a text file.

  • fast_atof, strtoul10, strtoul16, SkipSpaceAndLineEnd, SkipToNextToken .. large family of low-level parsing functions, mostly declared in fast_atof.h, StringComparison.h and ParsingUtils.h (a collection that grew historically, so don’t expect perfect organization).

  • ComputeNormalsWithSmoothingsGroups() (SmoothingGroups.h) - Computes normal vectors from plain old smoothing groups.

  • SkeletonMeshBuilder (SkeletonMeshBuilder.h) - generate a dummy mesh from a given (animation) skeleton.

  • StandardShapes (StandardShapes.h) - generate meshes for standard solids, such as platonic primitives, cylinders or spheres.

  • BatchLoader (BaseImporter.h) - manage imports from external files. Useful for file formats which spread their data across multiple files.

  • SceneCombiner (SceneCombiner.h) - exhaustive toolset to merge multiple scenes. Useful for file formats which spread their data across multiple files.

Filling materials

The required definitions zo set/remove/query keys in #ai_material structures are declared in MaterialSystem.h, in a #aiMaterial derivate called #aiMaterial. The header is included by AssimpPCH.h, so you don’t need to bother.

aiMaterial* mat = new aiMaterial();

const float spec = 16.f;
mat->AddProperty(&spec, 1, AI_MATKEY_SHININESS);

//set the name of the material:
NewMaterial->AddProperty(&aiString(MaterialName.c_str()), AI_MATKEY_NAME);//MaterialName is a std::string

//set the first diffuse texture
NewMaterial->AddProperty(&aiString(Texturename.c_str()), AI_MATKEY_TEXTURE(aiTextureType_DIFFUSE, 0));//again, Texturename is a std::string

Appendix A

Template for BaseImporter’s abstract methods

// -------------------------------------------------------------------------------
// Returns whether the class can handle the format of the given file.
bool xxxxImporter::CanRead( const std::string& pFile, IOSystem* pIOHandler,
      bool checkSig) const {
    const std::string extension = GetExtension(pFile);
    if(extension == "xxxx") {
        return true;
    if (!extension.length() || checkSig) {
        // no extension given, or we're called a second time because no
        // suitable loader was found yet. This means, we're trying to open
        // the file and look for and hints to identify the file format.
        // #Assimp::BaseImporter provides some utilities:
        // #Assimp::BaseImporter::SearchFileHeaderForToken - for text files.
        // It reads the first lines of the file and does a substring check
        // against a given list of 'magic' strings.
        // #Assimp::BaseImporter::CheckMagicToken - for binary files. It goes
        // to a particular offset in the file and and compares the next words
        // against a given list of 'magic' tokens.

        // These checks MUST be done (even if !checkSig) if the file extension
        // is not exclusive to your format. For example, .xml is very common
        // and (co)used by many formats.
    return false;

// -------------------------------------------------------------------------------
// Get list of file extensions handled by this loader
void xxxxImporter::GetExtensionList(std::set<std::string>& extensions) {

// -------------------------------------------------------------------------------
void xxxxImporter::InternReadFile( const std::string& pFile,
      aiScene* pScene, IOSystem* pIOHandler) {
    std::unique_ptr<IOStream> file( pIOHandler->Open( pFile, "rb"));

    // Check whether we can read from the file
    if( file.get() == NULL) {
        throw DeadlyImportError( "Failed to open xxxx file ", pFile, ".");

    // Your task: fill pScene
    // Throw a ImportErrorException with a meaningful (!) error message if
    // something goes wrong.