GND files contain the entirety of a map's static geometry (i.e., the ground/terrain). In the latest version they also include the configuration of the water surfaces, which were previously stored in the RSW file.

In order to understand the GND file format, some familiarity with the following concepts is required:

• The GAT specification, as it closely ties into how the GND file is interpreted
• Polygon meshes, used to represent objects in 3D space
• Lightmaps, a standard way to use precomputed lighting and improve rendering performance
• Posterization, an effect of reducing the number of colors which is applied to the lightmap texture

The ground mesh is a static object that represents the ground (or terrain) of each map. It is the defining part of the GND file and consists chiefly of geometry to which diffuse textures, color highlights and lightmap textures are applied.

Both files are defining different parts of a map's geometry. See also the following quote:

What you can take away from the RO method, is that they actually use two heightmaps: one invisible heightmap used for collision detection, and a visible one used to draw the map. That way, they can extend the "invisible wall or platforms" around various objects (eg those boats or docks placed in the scene that are not part of the heightmap

Source: Game Development @ StackExchange

The "collision map" is basically the contents of the GAT file (terrain types and "walkable" flags), while the "height map" is chiefly equivalent to the ground mesh as defined by its height vectors. The server needs to have the GAT information but doesn't care about GND, while the client uses the GAT to place units at their correct height and the GND to render the terrain geometry itself.

Recreating the terrain involves a complex set of operations; It is stored in a memory-efficient format and therefore not directly compatible with how modern rendering hardware expects 3D objects to be described (see Face-Vertex Mesh).

Instead of storing the vertex data directly, individual chunks of terrain ("cubes") are implicitly defined by the "height values" that are effectively defining the upwards-facing surface geometry by giving the relative position of its corners, offset by the surface cube's position in an imaginary "grid" of cubes.

To better explain how the geometry can be reconstructed, it might be helpful to view it as a combination of different (simpler) primitive shapes that are "stitched together" to form the ground mesh.

RO uses tiles as the basic unit of measurement for game logic, but its world actually consists of surfaces larger than that, which are part of what you could call "cubes" (boxes). These boxes are combined like a jigsaw puzzle to form the actual terrain.

Cubes are defined by only three surfaces (TOP, NORTH, EAST), which can be one of the two types mentioned below. In order to realize the BOTTOM, SOUTH and WEST sides, the adjacent cube's TOP, NORTH and EAST surface are used, respectively. The only two types of surfaces are GROUND and WALL, where GROUND is essentially the square defined by connecting the four height vectors stored in the GND file, and WALLS are the implied surfaces formed by connecting the GROUND surface of one cube to the GROUND surface of the neighboring cube, forming a wall where they differ in height.

Each surface of a cube is 2x2 GAT tiles large and may be textured or not. Surfaces without textures (i.e., texture ID is -1) aren't rendered, nor can they be used to attach "wall" surfaces to from their adjacent cube. This is relevant, because in at least two instances, namely c_tower4 and juperos_01, ground surfaces on the map boundaries are erroneously assigned EAST and NORTH surfaces, but you can't render walls on map boundaries since there's no vertices that you could connect with.

There's no direct reference to the tiles as defined in the world coordinate system, but rather surfaces/cubes are combined in a predefined order (bottom/left to top/right, with new rows starting whenever the "map width" has been exceeded - very much like pixels in a bitmap image).

Since this process is somewhat convoluted, more details are given below.

Also, here's an amazing elementary-school-level illustration of the above concepts:

Pictured: The height values from the GND file's cube grid define a set of height vectors for each cube.

These then have to be turned into an equivalent representation of the actual geometry:

Any surfaces and cubes that are implicitly defined by the height vectors can easily be reconstructed.

The green tile selector cursor here serves as a comparison for the dimensions; as mentioned above, each surface takes up the space of two GAT tiles. Wall surfaces may often appear bigger, but from what I can tell they're still mapped just the same and appear stretched if they happen to be larger.

As previously mentioned, tiles are the basic building block for the terrain, so in a way RO could be described as a tile-based game. It's not technically accurate, but that's definitely how the end result is perceived. Characters are positioned (at the center) on a tile, map coordinates refer to the tile and everything that requires thinking about positions usually works with the tile as the smallest unit... except when it comes to rendering.

In practical terms, the actual size of each tile, as rendered on the screen, depends on the scale of the world, camera zoom level and screen resolution. However, it's the baseline for measuring distance in the world coordinate system, where one unit of distance can be expressed in "number of tiles".

As far as textures are concerned, a texture appears to fit on a single tile if its dimensions are 32 pixels, as can easily be seen by checking the size of the "grid selector" texture rendered on top of the currently selected tile.

Surfaces are the visible areas containing tiles. The entirety of the map's geometry is made out of surfaces, which can be textured or "invisible" and as far as I can tell, each surface contains exactly four tiles (two in each dimension).

This is mostly of relevance for texturing, since texture slices usually are 64 pixels wide, so larger surfaces that appear seamless in the game are frequently pieced together by individual surfaces to give the appearance of a continuous textured plane (and they're indeed rendered similarly). Unfortunately this can cause problems if texture coordinates aren't applied correctly (and sometimes even if they are), which is why there are visible "gaps" in the terrain that can be seen at the right camera angle and distance.

The vertices forming the terrain's surface geometry must be connected from bottom-right to top-left for each textured surface. If done the other way around some slopes will look very wrong.

This beautiful part of early 2000s wall geometry can be found at approximately pay_dun00 (40, 35):

Connecting the vertices the wrong way (bottom-left to top-right) results in a visible glitch:

Clearly, the makers of this cave did not intend for things to go so horribly wrong!

Luckily, connecting them the other way around eliminates the problem:

For comparison, here's the original rendition once more:

So remember kids, always draw your triangles the right way... or someone will be very sad when they see the result.

Generally speaking, the terrain is defined by a heightmap indicating the positions of all four corners for any given tile. Each chunk of the heightmap (2x2 tiles) is sometimes called cube, which implies the presence of up to 6 surfaces: Two for the top and bottom of the cube, two to either side, and two to the front and back, where of course any number of them could be missing.

Now, the reason the cube metaphor doesn't work perfectly is that each chunk is only defined by three surfaces at most: Top, north, and east. If interpreted as surfaces it doesn't really change anything; Since you can express a cube's left (or south)-facing side as the adjacent cube's right (or north)-facing side you get the same effect with a much smaller data structure.

This visualization might help understand the implications. In the picture, surfaces are colored as follows:

• Green if they're the TOP surface for a given cube
• Blue if they're NORTH (wall surfaces)
• Red if they're EAST (wall surfaces)

What this means is that, counter to intuition, the blue parts to the south of each chunk aren't really the south surface of the respective ground cube, but rather the north one of the neighbouring cube to the south. The "walls" are implicitly defined by the difference in both cubes' height vector and the client renders them according to which of the top surface is higher, or not at all if the surface's aren't textured.

A side effect is that the northernmost "cubes" of each map can't have a "northern" wall, nor can the westernmost "cubes" have a wall to the west. This is why every single map has a specially-crafted "border" beyond the actual visible "walls" (if there are any), which are black impassable surface that are just sitting there and can't usually be seen in the game, as they blend in with the background.

In the GND files, you will merely find each corner's height and the overall scale factor, which is enough to render the geometry. You could do this by translating the positions to world coordinates and normalize the heights so they refer to a cube of width and height 1 (as in, "one unit in world coordinates"), or just make the cubes bigger to account for the scale factor and then "zoom out" farther to get the same perspective.

The geometry is then rendered by simply glueing all the surfaces together, and walls are implicitly formed by two tiles that aren't on the same height: If there's a height difference there'll be a plane in between the two tiles; this essentially becomes a "wall" if assigning it a texture. However, this is only done if the neighbouring "cube" is higher or lower and has a textured surface of its own.

Each entity in the game world is assigned a "zoom" factor that is used to calculate its rendered size on the screen. For objects of type "Ground", this is always initialized with the value 10. Starting at version 1.6, it is overwritten with whatever value is stored in the map's GND file.

Since I haven't seen any map that used a different "zoom level" for its terrain, I have to assume they just kept the default value.

The actual calculation appears to be as follows:

normalizingScaleFactor = GND.mapDimensions / GAT.mapDimensions * GND.geometryScaleFactor

Since each GND surface normally measures two GAT tiles, and the scale factor is seemingly always set to ten, this nets:

normalizingScaleFactor = 1/2 * 10 = 1/5

Therefore a constant scale factor of 1/5 should be correct for every map. Applying this factor scales the heights given by the height vectors of the GND file back to the standard coordinate system, with one tile being one world unit large.

A lightmap is used to add static light and shadow detail to each map, without requiring the client to calculate lighting in real time.

Similarly to the terrain's surface itself, the lightmap texture is sliced into unique parts, and each part must be mapped to the right surface on the assembled ground mesh. Here's an example:

One possible rendition of the lightmap for c_tower2.gnd

Here, you can clearly see the individual slices that are mapped to each surface. The visible borders constitute the buffer areas intended to minimize artifacts from texture bleeding.

The ambient occlusion values are stored in the same format, with "white" pixels representing areas where no ambient occlusion (i.e., no "shadows") should be rendered. In this example, the buffer areas are even more starkly visible:

One possible rendition of the ambient occlusion map for c_tower2.gnd

The process for mapping slices to surfaces is otherwise similar to how the actual terrain is formed: Each surface references a slice of the lightmap texture, as well as a part of the regular (diffuse) texture, while each cube references up to three surfaces.

With GNDedit, it's trivial to export a human-readable form of the same textures:

Surface-mapped lightmap texture for c_tower2.gnd

The ambient occlusion map is derived by separating the alpha channel:

Surface-mapped ambient occlusion texture for c_tower2.gnd

These images only include the GROUND surfaces (walkable area). The process for mapping lightmap slices to WALL surfaces is identical, although it's slightly more difficult to visualize in this form.

The visible white borders in that screenshot bring us to another phenomenon: Texture bleeding. If all the tiny slices were to be combined seamlessly, there could be visible artifacts from texture bleeding as soon as linear filtering is enabled.

To avoid this problem, there's a one-pixel area at the outside of each texture slice that isn't used directly. It serves to provide a buffer between the adjacent slices and removes or at least reduces the effect when texture filtering is applied.

In this image I've manually merged two adjacent slices:

The green area marks the visible portion of each slice that is contained within the area defined by the texture coordinates, with the area outside being the "buffer" that is blended with its adjacent texels when texture bleeding occurs.

Lastly, there isn't a single lightmap texture, but rather two: A "color map" that contains static lighting, and a "shadow map" that contains only greyscale data and is used to render shadows. The latter effectively stores the ambient occlusion percentage for each pixel.

Example: Color map vs Shadow map

The "shadow map" must be applied after the lighting calculation to avoid having it blended in with the other light sources, which would cause shadows to appear "washed out" or even invisible, depending on when in the lighting process the colors are multiplied with the shadow/alpha values.

I believe the client might simply use the alpha channel directly and store both in just one texture, since there's just one texture stored in the GND file and the /lightmap command seems to disable both parts, but this is merely an educated guess.

The client reduces the number of colors when processing the lightmap texture, causing visible artifacts. These are not part of the lightmap itself, but rather introduced deliberately when loading the texture.

I can only speculate as to why such a "feature" exists, but my guess would be technical limitations or a stylistic choice. Either way, it is possible to render lighting in the way it was originally computed using just the information that's part of the GND file. This smoothes the transition between colors, as can be seen in the following comparison:

• Here are the lights of Comodo in their original glory
• And here they are again after applying the posterization effect with a "posterization level" (see below) of 16

Note the jagged appearance that plays a huge part in giving RO maps their distinctive look.

The effect is achieved by applying the following operation to the RGB values:

r = (int)(r / LEVEL_COUNT) * LEVEL_COUNT
g = (int)(g / LEVEL_COUNT) * LEVEL_COUNT
b = (int)(b / LEVEL_COUNT) * LEVEL_COUNT
In RO, LEVEL_COUNT is 16.

For javascript it would be c = floor(c / LEVEL_COUNT) * LEVEL_COUNT;
where c is each color component

Source: greenboxal

This effectively casts the color value from float to int, reducing the amount of visible colors and thereby introducing the "jagged" transitions in the color gradient.

In order to give highlights to the terrain without requiring vast amounts of slightly different textures, the client can render specific (solid) colors at the corners of each tile to give this corner (and all adjacent ones) a different hue.

This concept is known as vertex colors and it works by submitting additional information for a given point in 3D space that is then used by the hardware/GPU software to color it differently. It was widely used in older games where dynamic lighting wasn't feasible due to technical limitations, and it can have a significant visual impact for barely any effort at all:

Again they massively saved on file size by providing only one color value per tile, which is applied to the bottom left corner (vertex) of each tile and the vertices of all adjacent tiles that happen to be in the same position. With this, one can define colored spots on the corners of the tile grid with very little effort or overhead.

In this basic rendition of pay_dun00 you can see the textured terrain with vertex colors applied (lightmaps, models, etc. are not displayed). They're the "blotches" on the ground, which blend in with the rest of the scene after all other effects are applied and give the terrain a lot of variety while still using a very limited set of textures.

Each tile can have a diffuse color. This color, however, is not being applied to all four corners of a tile, but only to the bottom left vertex and all vertices that share the same coordinate.

At least one version exists that doesn't feature the typical GND header.

It seems to otherwise be largely identical to the later version 1.7, but I think it may use a different (simpler) lightmap format.

This is only an educated guess based on comparing the converted legacy GND files (author unknown) published on the athena forums.

This is the original version used in the modern RO client.

It contains all the basic information about the map's surface geometry, as well as information about the diffuse textures and lightmap/ambient occlusion textures.

There are two textures encoded in the files, one is a prebaked lightmap and the other a "shadowmap" (ambient occlusion map). They are read in that same order, based on the lightmap pixel format.

These are just the blueprints for creating the terrain geometry, and have to be applied (by copying over the texturing information) to the actual vertices implied by the ground mesh cubes.

The texture coordinates are stored in order

1. bottomLeftU
2. bottomRightU
3. topLeftU
4. topRightU
5. bottomLeftV
6. bottomRightV
7. topLeftV
8. topRightV

where the direction is always relative to the cube that the surface belongs to.

Please note that the dimensions are implied by the "width" and "height" given at the start of the file, and that they are different (half as big) as the actual map width and height given in GAT tiles. This is because each GND cube surface is 2 GAT tiles large, as described above.

The water plane configuration that was previously part of the RSW file has been removed from the RSW files and is now stored in the map's GND file instead.

All the information stored in version 1.7 is still present, plus the water plane configuration (appended).

In theory, the new GND format should support multiple water planes. In reality, all the maps that use it only feature one and are therefore indistinguishable from those using the old RSW format.

The reason is that a new version was introduced quickly afterwards, replacing 1.8 with a more elaborate version 1.9 that also supports this feature.

This is the basic water configuration, as present in both RSW (pre-2.6) and the next GND version. The configuration for the other planes follows after the above, and is (in this version) always of the following form:

1. numWaterPlanesU (int32): How many water planes there are in the horizontal dimension (always 1)
2. numWaterPlanesV (int32): How many water planes there are in the vertical dimension (always 1)
3. A float32 defining the waterLevel (altitude) of this plane. It's always identical to the water level above, since there is only one.

If there were multiple values here, it would likely be the water level of each plane (assuming the presence of u * v planes), but since that doesn't happen and there is already a new version it seems unlikely that this would ever be used.

All the information stored in version 1.7, plus the water plane configuration (appended). The latter consists of the same information as is present in 1.8, but the format for defining multiple water planes has changed (and they are in fact used in some of the episode 19 maps I've seen, such as icecastle.gnd):

1. numWaterPlanesU (int32): How many water planes there are in the horizontal dimension (not always 1)
2. numWaterPlanesV (int32): How many water planes there are in the vertical dimension (not always 1)
3. The same structure as defined for the original water plane (see table above), repeated u * v times

I'm not quite sure what the original configuration is still used for, but I'm guessing it would serve as a fallback in case the numbers (u, v) are both zero, i.e. there aren't any water planes defined in this section.

The individual water planes are placed in a grid of u * v cells, ordered just like the ground mesh cubes from bottom-left to top-right (in map coordinates). Each plane is exactly mapWidth / u wide and mapHeight / v high (all given in GAT coordinates), in combination covering the entire map. This is different from the situation where there's just a single plane, which would instead cover the entire map on its own.

• This online demo showing GND-style cube geometry defined in a standardized vertex data format (Note: The triangles are formed bottom-right to top-left in the actual GND ground mesh and not bottom-left to top-right, as displayed here)
• An intuitive explanation of the terrain in RO (archived; all screenshots are missing)