Voxel block format v2


This document is about an old version of the format. You may check the most recent version.

Version: 2

This page describes the binary format used by default in this module to serialize voxel blocks to files, network or databases.

Changes from version 1

Overall the format can be more standalone than before, where information had to be known up-front.

  • Made compressed container a bit more independent with a header
  • Added version header
  • Added 3D size
  • Added depth information on each channel

There is no migration available from version 1.


Compressed container

A block is usually serialized as compressed data. This is the format provided by the VoxelBlockSerializer utility class. If you don't use compression, the layout will correspond to BlockData described in the next listing, and won't have this wrapper.

Compressed data starts with one byte. Depending on its value, what follows is different.

  • 0: no compression. Following bytes can be read as as block format directly. This is rarely used and could be for debugging.
  • 1: LZ4 compression. The next big-endian 32-bit unsigned integer is the size of the decompressed data, and following bytes are compressed data using LZ4 default parameters. This mode is used by default.

Knowing the size of the decompressed data may be important when parsing the block later.

Block format

The obtained data then contains the actual block.

It starts with version number 2 in one byte, then some metadata and the actual voxels.


The size and formats are present to make the format standalone. When used within a chunked container like region files, it is recommended to check if they match the format expected for the volume as a whole.

- version: uint8_t
- size_x: uint16_t
- size_y: uint16_t
- size_z: uint16_t
- channels[8]
- metadata*
- epilogue


Block data starts with exactly 8 channels one after the other, each with the following structure:

- format: uint8_t (low nibble = compression, high nibble = depth)
- data

format contains both compression and bit depth, respectively known as VoxelBuffer::Compression and VoxelBuffer::Depth enums. The low nibble contains compression, and the high nibble contains depth. Depending on those values, data will be different.

Depth can be 0 (8-bit), 1 (16-bit), 2 (32-bit) or 3 (64-bit).

If compression is COMPRESSION_NONE (0), data will be an array of N*S bytes, where N is the number of voxels inside a block, multiplied by the number of bytes corresponding to the bit depth. For example, a block of size 16x16x16 and a channel of 32-bit depth will have 16*16*16*4 bytes to load from the file into this channel. The 3D indexing of that data is in order ZXY.

If compression is COMPRESSION_UNIFORM (1), the data will be a single voxel value, which means all voxels in the block have that same value. Unused channels will always use this mode. The value spans the same number of bytes defined by the depth.

Other compression values are invalid.


After all channels information, block data can contain metadata information. Blocks that don't contain any will only have a fixed amount of bytes left (from the epilogue) before reaching the size of the total data to read. If there is more, the block contains metadata.

- metadata_size: uint32_t
- block_metadata
- voxel_metadata[*]

It starts with one 32-bit unsigned integer representing the total size of all metadata there is to read. That data comes in two groups: one for the whole block, and one per voxel.

Block metadata is one Godot Variant, encoded using the encode_variant method of the engine.

Voxel metadata immediately follows. It is a sequence of the following data structures, which must be read until a total of metadata_size bytes have been read from the beginning:

- x: uint16_t
- y: uint16_t
- z: uint16_t
- data

x, y and z indicate which voxel the data corresponds. data is also a Variant encoded the same way as described earlier. This results in an associative collection between voxel positions relative to the block and their corresponding metadata.


At the very end, block data finishes with a sequence of 4 bytes, which once read into a uint32_t integer must match the value 0x900df00d. If that condition isn't fulfilled, the block must be assumed corrupted.


On little-endian architectures (mostly desktop), binary editors will not show the epilogue as 0x900df00d, but as 0x0df00d90 instead.

Current Issues

Although this format is currently implemented and usable, it has known issues.


Godot's encode_variant doesn't seem to care about endianness across architectures, so it's possible it becomes a problem in the future and gets changed to a custom format. The rest of this spec is not affected by this and assumes we use little-endian, however the implementation of block channels with depth greater than 8-bit currently doesn't consider this either. This might be refined in a later iteration.

This will become important to address if voxel games require communication between mobile and desktop.