Dev:Source/Architecture/File Format

提供: wiki
移動先: 案内検索

The blend-file-format


This document is based on the document The Mystery of the Blend written by Jeroen Bakker, slightly edited. The article describes the basics of the blend-file-format by stepping over the following topics:

  • First we describe how Blender works with blend-files.
  • We look at the global file-structure of a blend-file (the file-header and file-blocks).
  • After this is explained, we go deeper to the core of the blend-file, the DNA-structures. They hold the blue-prints of the blend-file and the key asset of understanding blend-files.
  • When that's done we can use these DNA-structures to read information from elsewhere in the blend-file.

As a working example we'll be using the default blend-file from Blender 2.48, with the goal to read the output resolution from the Scene. The article is written to be programming language independent.

Loading and saving in Blender

Saving and loading complex scenes in Blender is done within seconds. Blender achieves this by saving data in memory to disk without any transformations or translations. Blender only adds file-block-headers to this data.

A file-block-header contains clues on how to interpret the data. After the data, all internally Blender structures are stored. These structures act as blue-prints when Blender loads the file.

Blender3D FreeTip.png
Blend File Format is self describing
Blend-files can be different when stored on different hardware platforms or Blender releases. There is no effort taken to make blend-files binary the same. Backward and Forward compatibility is maintained during loading a blend file into Blender.

When Blender loads a blend-file, it first creates a catalog of DNA-structures (see further down for what DNA structures are). This catalog is then used to ensure the backward and forward compatibility of the loaded data with Blender's internal structures. Within the Blender sourcecode you find the logic for transforming every structure used by a Blender release to the one of the release you're using, see /source/blender/blenloader/intern/readfile.c. The more difference between releases the more logic is executed.

The blend-file-format does not differ from internally used structures and the file is self contained and self explaining (after you read this article).

Global file-structure

A blend-file always start with the file-header followed by one or more file-blocks. Each file-block has a file-block-header and data.

Sidenote: the default blend file of Blender 2.48 contains more than 400 file-blocks.


The first 12 bytes of every blend-file is the file-header. The file-header has information on Blender (version-number) and the PC the blend-file was saved on (pointer-size and endianness). This is required as all data inside the blend-file is ordered in that way, because no translation or transformation is done during saving. The next table describes the information in the file-header:

reference structure type offset size
identifier char[7] File identifier (always 'BLENDER') 0 7
pointer-size char Size of a pointer

All pointers in the file are stored in this format
'_' (underscore) means 4 bytes or 32 bit
'-' (minus) means 8 bytes or 64 bits.

7 1
endianness(*) char Type of byte ordering used

'v' means little endian
'V' means big endian.

8 1
version-number char[3] Version of Blender the file was created in

for Example '248' means version 2.48

9 3

(*) The Endianness addresses the way values are ordered in a sequence of bytes [ref:]. Blender supports little-endian and big-endian. In a big endian ordering, the largest part of the value is placed on the first byte and the lowest part of the value is placed on the last byte. In a little endian ordering, largest part of the value is placed on the last byte and the smallest part of the value is placed on the first byte. Example: writing the integer 0x4A3B2C1Dh, will be ordered in Big endian as 0x4Ah, 0x3Bh, 0x2Ch, 0x1Dh and be ordered in little endian as 0x1Dh, 0x2Ch, 0x3Bh, 0x4Ah.

Blender3D FreeTip.png
The endianness can be different between the blend-file and the PC your using
When these are different, Blender changes it to the byte ordering of your PC. Nowadays, little-endian is the most commonly used.

The next hex-dump describes a file-header created with blender 2.48 on little-endian hardware with a 32 bits pointer length:

0000 0000: [42 4C 45 4E  44 45 52] [5F] [76] [32 34 38] (binary data)
0000 0000    B  L  E  N   D  E  R    _    v    2  4  8  (readable form)


File-blocks contain a file-block-header and data. The start of a file-block is always aligned at a 4 byte voundary. The file-block-header describes the total length of the data, the type of information stored in the file-block, the number of items of this information and the old memory pointer at the moment the data was written to disk.

Depending on the pointer-size stored in the file-header, a file-block-header can be 20 or 24 bytes long. The next table describes how a file-block-header is structured:

reference structure type offset size
code char[4] Identifier of the file-block 0 4
size integer Total length of the data
after the file-block-header
4 4
old memory address void* Memory address
pointer to where the structure
was located when written to disk
8 pointer-size (4/8)
SDNA index integer Index of the SDNA structure 8+pointer-size 4
count integer Number of structure located
in this file-block
12+pointer-size 4

code: describes different types of file-blocks. The code determines with what logic the data must be read. These codes also allows fast finding of data like Library, Scenes, Object or Materials as they all have a specific code.

size: contains the total length of data after the file-block-header. After the data a new file-block starts. The last file-block in the file has code 'ENDB'.

old memory address: contains the memory address when the structure was last stored. When loading the file the structures can be placed on different memory addresses. Blender updates pointers to these structures to the new memory addresses.

SDNA index: contains the index in the DNA structures to be used when reading this file-block-data. More information about this subject will be explained in the Reading scene information section.

Count: tells how many elements of the specific SDNA structure can be found in the data.

The next section is an example of a file-block-header.

           code:        size:        old mem:     SDNA:
0000 4420: 53 43 00 00  60 05 00 00  A0 2F 04 0A  8B 00 00 00 (binary data)
            S  C  .  .   `  .  .  .   .  /  .  .   .  .  .  . (readable form)

0000 4430: 01 00 00 00  xx xx xx xx  xx xx xx xx  xx xx xx xx (binary data)
            .  .  .  .   x  x  x  x   x  x  x  x   x  x  x  x (readable form)
  • code (first 4 bytes) 'SC'+0x00h identifies that it is a Scene.
  • Size (next 4 bytes big endian): 0x60h + 0x05h * 256 = 96 + 1280 = 1376 bytes
  • old pointer is: 0x0A042FA0h
  • SDNA index is : 0X8B = 8*16 + 11 = 139.
  • count: 1 in this example means: the section contains a single scene.

Before we can interpreted the data of this file-block we first have to read the DNA structures in the file. The section structure DNA will show how to do that:

Structure DNA

Structure DNA is stored in a file-block with code 'DNA1'. It can be just before the 'ENDB' file-block. It contains all internal structures of the Blender release the file was created in. The data in this file-block must be interpreted as described in the this section.

In a blend-file created with Blender 2.48a this section is 43468 bytes long and contains 309 structures. These structures can be described as C-structures. They can hold fields, arrays and pointers to other structures, just like a normal C-structure:

structure Scene {
    ID id;           // 52 bytes (ID is different a structure)
    Object *camera;  //  4 bytes (pointer to an Object structure)
    World *world;    //  4 bytes (pointer to a World structure)
    float cursor[3]; // 12 bytes (array of 3 floats)

The next section describes how this information is ordered in the data of the 'DNA1' file-block.

repeat condition name type length description
identifier char[4] 4 'SDNA'
name identifier char[4] 4 'NAME'
#names integer 4 Number of names follows
for(#names) name char[] ? Zero terminated string of name
may contain pointer and
simple array definitions
e.g.: '*vertex[3]\0'
type identifier char[4] 4 'TYPE' (this field is aligned at 4 bytes)
#types integer 4 Number of types follows
for(#types) type char[] ? Zero terminated string of type (e.g. 'int\0')
length identifier char[4] 4 'TLEN' (this field is aligned at 4 bytes)
for(#types) length short 2 Length in bytes of type (e.g. 4)
structure identifier char[4] 4 'STRC' (this field is aligned at 4 bytes)
#structures integer 4 Number of structures follows
for(#structures) structure type short 2 Index in types containing the name of the structure
.. #fields short 2 Number of fields in this structure
.. for(#field) field type short 2 Index in type
for end for end field name short 2 Index in name

As you can see, the structures are stored in 4 arrays: names, types, lengths and structures.

Each structure also contains an array of fields. A field is the combination of a type and a name. From this information a catalog of all structures can be constructed.

The names are stored as how a C-developer defines them. This means that the name also defines pointers and arrays:

  • If a name starts with '*' it is used as a pointer.
  • If the name contains for example '[3]' it is used as a array of length 3.

In the types you'll find simple-types (like: 'integer', 'char', 'float'), but also complex-types like 'Scene' and 'MetaBall'. 'TLEN' part describes the length of the types. A 'char' is 1 byte, an 'integer' is 4 bytes and a 'Scene' is 1376 bytes long.

While reading the DNA you'll will come across some strange names like '(*doit)()'. These are method pointers and Blender updates them to the correct methods.

The fields 'type identifier', 'length identifier' and 'structure identifier' are aligned at 4 bytes.

The DNA structures inside a Blender 2.48 blend-file can be found at If we understand the DNA part of the file it is now possible to read information from other parts file-blocks. The next section will tell us how.

Reading scene information

Let us look at the file-block we have seen earlier. The code is 'SC'+0x00h and the SDNA index is 139. The 139th structure in the DNA is a structure of type 'Scene'. The associated type ('Scene') has the length of 1376 bytes. This is exact the same length as the data in the file-block. We can map the Scene-structure on the data of the file-blocks. But before we can do that, we have to flatten the Scene-structure.

The first field in the Scene-structure is of type 'ID' with the name 'id'. Inside the list of DNA structures there is a structure defined for type 'ID' (structure index 17). The first field in this structure has type 'void' and name '*next'. Looking in the structure list there is no structure defined for type 'void'. It is a simple type and therefore the data should be read. '*next' describes a pointer. the first 4 bytes of the data can be mapped to ''. Using this method we'll map a structure to its data. If we want to read a specific field we know at what offset in the data it is located and how much space it takes.

The next table shows the output of this flattening process for some parts of the Scene-structure. Not all rows are described in the table as there is a lot of information in a Scene-structure. flattened SDNA structure 139: Scene

reference structure type name offset size description ID void *next 0 4 Refers to the next scene
id.prev ID void *prev 4 4 Refers to the previous scene
id.newid ID ID *newid 8 4
id.lib ID Library *lib 12 4 ID char name[24] 16 24 'SC'+the name of the scene as displayed in Blender ID short us 40 2
id.flag ID short flag 42 2
id.icon_id ID int icon_id 44 4 ID IDProperty *properties 48 4
camera Scene Object *camera 52 4 Pointer to the current camera
world Scene World *world 56 4 Pointer to the current world
set Scene Scene *set 60 4 Pointer to the current set
Skipped rows
r.sfra RenderData int sfra 248 4 Start frame of the scene
r.efra RenderData int efra 252 4 End frame of the scene
Skipped rows
r.xsch RenderData short xsch 326 2 X-resolution of the output when rendered at 100%
r.ysch RenderData short ysch 328 2 Y-resolution of the output when rendered at 100%
r.xparts RenderData short xparts 330 2 Number of x-part the renderer uses
r.yparts RenderData short yparts 332 2 Number of y-part the renderer uses
Skipped rows
sculptdata.axislock SculptData char axislock 1365 1
sculptdata.pad SculptData char pad[2] 1366 2
frame_step Scene int frame_step 1368 4
pad Scene int pad 1372 4

We can now read the X and Y resolution of the Scene. The X-resolution is located on offset 326 of the file-block-data and must be read as a short. The Y-resolution is located on offset 328 and is also a short.

An array of chars can mean 2 things. The field contains readable text or it contains an array of flags (not humanly readable).
A file-block containing a list refers to the DNA structure and has a count larger than 1. For example Vertexes and Faces are stored in this way.

Next steps

The implementation of saving in Blender is easy, but loading is difficult. When implementing loading and saving blend-files in a custom tool the difficulty is the opposite. In a custom tool loading a blend-file is easy, and saving a blend-file is difficult. If you want to save blend-files I suggest to start with understanding the the global file structure and parsing the DNA section of the file. After this is done it should be easy to read information from existing blend files like scene data, materials and meshes. When you feel familiar with this you can start creating blend-libraries using the internal Blender structures of a specific release. If you don't want to dive into the Blender source code you can find them all at

There is a feature request on supporting an XML based import/export system in Blender. I dont support the request, but it is interesting to look at how this can be implemented. An XML export can be implemented with low effort as an XSD can be used as DNA structures and the data can be written into XML [see to download JAVA example including source code]. Implementing an XML import system uses a lot of memory and CPU. If you really want to implement it, I expect that the easiest way is to convert the XML-file back to a normal blend-file and then load it using the current implementation. One real drawback is that parsing a XML based blend-file uses a lot of memory and CPU and the files can become very large.

At this moment I'm using this information in an automated render pipeline. The render pipeline is build around a web-server and SVN. When an artist commits a new blend-file in SVN, it is picked up by the web-server and it will extract resolutions, frames scenes and libraries from the blend-file. This information is matched with the other files in SVN and the blend-file will be placed in the render pipeline.