Warcraft III MDX Model Format

MDX (also written as MDLX) is Warcraft III's binary 3D model format. It stores geometry, materials, bones, animations, particle emitters, and more in a chunked binary layout.

File Layout

An MDX file is a sequence of tagged chunks. Each chunk begins with a 4-byte FourCC identifier followed by a 4-byte chunk size in bytes. The top-level structure is:

MDLX                       ← magic / version header
VERS  <size>  <version>    ← format version (800 for WC3, 1000 / 1500 for Reforged)
MODL  <size>  <modelInfo>  ← global model info (name, bounds)
SEQS  <size>  [sequence]…  ← named animation sequences
GLBS  <size>  [globalSeq]… ← global sequence durations
TEXS  <size>  [texture]…   ← texture path list
MTLS  <size>  [material]…  ← materials (layer stacks)
GEOS  <size>  [geoset]…    ← geometry (vertices + faces)
BONE  <size>  [bone]…      ← skeleton bones
HELP  <size>  [helper]…    ← helper nodes (invisible pivot points)
ATCH  <size>  [attach]…    ← attachment points (e.g. "Overhead")
PIVT  <size>  [pivot]…     ← per-node pivot positions
PREM  <size>  [emitter]…   ← particle emitter v1
PRE2  <size>  [emitter2]…  ← particle emitter v2
RIBB  <size>  [ribbon]…    ← ribbon emitters
EVTS  <size>  [event]…     ← event objects (sounds, etc.)
CLID  <size>  [collision]… ← collision shapes
LITE  <size>  [light]…     ← light nodes

Not all chunks are present in every model. The loader (renderer/mdx/r_mdx_load.c) dispatches on each FourCC tag.

Node Hierarchy

All nodes (bones, helpers, attachments, emitters, event objects, collision shapes, lights) share a common node header:

Field	Type	Description
`size`	`DWORD`	Total size of this node record in bytes
`name`	`char[80]`	Human-readable name
`objectId`	`DWORD`	Unique node index (0-based)
`parentId`	`DWORD`	Parent node index (`0xFFFFFFFF` = root)
`flags`	`DWORD`	Node type and billboarding flags

The flags field uses the following bitmasks:

Flag	Value	Meaning
`Helper`	`0`	Plain helper (pivot only)
`DontInheritTranslation`	`1`	Ignore parent translation
`DontInheritRotation`	`2`	Ignore parent rotation
`DontInheritScaling`	`4`	Ignore parent scale
`Billboarded`	`8`	Always faces the camera
`BillboardedLockX`	`16`	Billboarded on X axis only
`BillboardedLockY`	`32`	Billboarded on Y axis only
`BillboardedLockZ`	`64`	Billboarded on Z axis only
`CameraAnchored`	`128`	Positioned relative to the camera
`Bone`	`256`	This node is a skeleton bone
`Light`	`512`	This node is a light source
`EventObject`	`1024`	Triggers events (sounds, effects)
`Attachment`	`2048`	Named attachment point
`ParticleEmitter`	`4096`	Particle emitter
`CollisionShape`	`8192`	Physics collision volume
`RibbonEmitter`	`16384`	Ribbon/trail emitter

Keyframe Tracks (Animated Values)

Animation data is stored as keyframe tracks. Each track is identified by a 4-byte tag that encodes the node type and the animated property. For example:

Tag	Target	Property
`KGTR`	Node	Translation
`KGRT`	Node	Rotation (quaternion)
`KGSC`	Node	Scale
`KMTA`	Material layer	Alpha
`KMTE`	Material layer	Emissive gain
`KP2V`	Particle emitter v2	Visibility
`KP2E`	Particle emitter v2	Emission rate
`KLAV`	Light	Ambient intensity

A keyframe track record starts with:

DWORD  numKeys
DWORD  interpolationType   // 0=none, 1=linear, 2=hermite, 3=bezier
DWORD  globalSeqId         // 0xFFFFFFFF if not driven by a global seq
[key × numKeys]

Each key is { DWORD frame; <value>; [<inTan>; <outTan>] } where the tangent pair is only present for hermite/bezier interpolation.

Geoset (Geometry)

A geoset is one draw call: a set of vertices sharing the same material. Its sub-chunks are:

Sub-chunk	Description
`VRTX`	Vertex positions — `float[3]` each
`NRMS`	Vertex normals — `float[3]` each
`UVBS`	UV sets — `float[2]` each
`PTYP`	Primitive types (always `4` = triangles)
`PCNT`	Primitive counts
`PVTX`	Vertex index list (triangle list)
`GNDX`	Bone group index per vertex
`MTGC`	Vertex count per bone group
`MATS`	Bone indices for each bone group

The renderer builds a static VBO from VRTX/NRMS/UVBS and a matrix palette from the MATS/GNDX/MTGC tables for GPU skinning.

Geoset Flags

The flags field of a geoset (mdxGeoFlags_t) controls render state:

Flag	Value	Meaning
`Unshaded`	`0x01`	No lighting — use vertex colour directly
`TwoSided`	`0x10`	Disable back-face culling
`Unfogged`	`0x20`	Not affected by distance fog
`NoDepthTest`	`0x40`	Always drawn on top
`NoDepthSet`	`0x80`	Do not write to the depth buffer

Material Layers

Each material is a stack of one or more layers rendered in order (additive blending for particle effects, etc.). A layer record includes:

filter mode (none, transparent, blend, additive, add-alpha, modulate, modulate2×)
texture index (into the TEXS table)
texture animation index
flag bits (unshaded, sphere env map, two-sided, unfogged, no-depth-test, no-depth-set)
animated alpha (KMTA track)

Animation Sequences

Each entry in SEQS describes one named clip:

Field	Type	Description
`name`	`char[80]`	Sequence name (e.g. `"Stand"`, `"Walk"`, `"Attack"`)
`interval`	`DWORD[2]`	[startFrame, endFrame] in milliseconds
`moveSpeed`	`float`	Ground speed during the animation
`flags`	`DWORD`	`1` = non-looping
`rarity`	`float`	Random weight for stand variations
`syncPoint`	`DWORD`	Sync reference frame
`extent`	`Extent`	Bounding box + sphere for this clip

Standard sequence names are: Stand, Walk, Attack, Attack Slam, Attack 2, Decay Flesh, Decay Bone, Death, Dissipate, Portrait, Spell, Stand Channel, Stand Ready, Stand Work.

Source	Purpose
`renderer/mdx/r_mdx.h`	All MDX struct definitions
`renderer/mdx/r_mdx_load.c`	Chunk parser and model loader
`renderer/mdx/r_mdx_render.c`	Per-frame skinning and draw calls
`renderer/mdx/r_mdx_interpolation.c`	Keyframe track evaluation