Notes on Procedural World Generation of Ubisoft’s Far Cry 5

procedural-generation
game-dev
notes
In this presentation from 2018, Etienne Carrier details the procedural pipeline developed for Far Cry 5 using Houdini and Houdini Engine, focusing on its objectives, tools, user workflow, underlying mechanics, and lessons learned during development.
Author

Christian Mills

Published

December 9, 2021

Source Material

Introduction and Challenges

  • Etienne Carrier, Technical Artist at Ubisoft Montreal (3 years), presents the procedural pipeline developed for Far Cry 5.
  • Challenge: Terrain undergoing constant changes during the project’s 2.5-year development.
    • Manual content placement (e.g., forests) becomes incoherent with each terrain iteration.
    • Repainting content manually after each terrain change is tedious and difficult to maintain consistency across a large world with multiple users.
    • Locking terrain early is unrealistic as iterations are crucial for game quality.

Objectives of the Procedural Pipeline

  • Develop a macro management tool to:
    • Fill the world with natural-looking content.
    • Ensure content consistency with terrain topology, adapting to changes (e.g., forest distribution remaining coherent despite terrain adjustments).
  • Automation:
    • Utilize Houdini, Houdini Engine, and nightly builds on build machines to fully refresh the world daily.
    • Build machines process different map sections, ensuring up-to-date world data each morning.
  • Deterministic Generation:
    • Ensure the generation yields identical results with the same inputs, regardless of the build machine used.
    • This is crucial for seamless map junctions and nav mesh generation.
  • User-Friendliness:
    • Provide in-editor tools for on-demand procedural generation alongside the nightly builds.

Procedural Tools Developed

  • Expanded beyond the initial mandate of biome distribution to include:
    • Freshwater Tools: Generate lakes, rivers, streams, and waterfalls.
    • Fence and Power Line Tools: Create fences and power lines along splines.
    • Cliff Generation: Generates cliffs on steep terrain.
    • Biome Tool: Spawns vegetation throughout the world.
    • Fog Density Map Generation: Creates a 2D map based on terrain topology and content placement to influence the fog shader.
    • Wall Map Terrain Generation: Generates a low-detail terrain representation for the world map, including miniature trees.

User Workflow: Filling an Empty Map

  1. Terraforming:
    • Initial world machine pass followed by artist-driven terraforming using in-editor tools.
  2. Freshwater Network:
    • Artists lay down a network of curves and splines to define rivers and other water bodies.
    • Procedural generation creates the water surface, waterside assets, and terrain texturing based on spline parameters (e.g., width).
  3. Cliff Generation:
    • Tool automatically generates cliffs based on terrain slope, with minimal user input.
    • Includes features like exclusion masks for finer control.
    • Often handled by nightly builds, relieving artists from frequent manual generation.
  4. Vegetation (Biome Painter):
    • Artists paint sub-biomes (e.g., forest) using a painter tool.
    • Main biomes automatically distribute vegetation (grass, trees) based on terrain data and sub-biome information.
    • Recipes react to water proximity, altitude, and cliff erosion lines, spawning appropriate assets.
  5. Customization:
    • Artists can override the natural distribution by clearing areas with grass sub-biome and adding elements like roads (using splines).
    • Road generation clears vegetation and adds roads, requiring some manual asset placement (houses, vehicles).
    • Biome painting can be further refined to create features like driveways using terrain textures.
    • Forest sub-biome can be repainted to add trees around specific locations.
  6. Non-Destructive Workflow:
    • Terrain can be adjusted at any time, requiring a refresh of affected procedural tools (e.g., cliffs, biomes).
    • System maintains coherence and updates content placement automatically.
  7. Fence and Power Line Placement:
    • Fences are generated along splines with fence type specified as an attribute.
    • Power lines connect electric poles placed at spline control points.
    • System supports multiple power line types and automatically handles snapping and transformer placement.
    • Biome tool automatically clears vegetation obstructing power lines.

Under the Hood: Houdini Engine and Data Exchange

  • Houdini Engine within Dunia enables seamless data exchange between Houdini and the engine.
  • Inputs from Dunia to Houdini (via Python script):
    • World information (name, size).
    • File paths for assets and data.
    • Terrain sector information (area to generate).
    • Splines and shapes with metadata on geometry attributes (e.g., fence type on a fence spline).
  • Inputs from Disk (file paths provided by Dunia):
    • Height maps.
    • Biome painter data.
    • SPNG and 2D terrain masks.
    • Houdini geometry from previously generated tools.
  • Primary Input: The terrain itself, as generation is linked to specific terrain areas (sectors).
  • Terrain Subdivision:
    • Sectors are the smallest unit (64x64 meters), defining the minimum area for procedural generation.
  • Baking Procedural Data:
    • Users select a generation area:
      • All (everything loaded in the editor).
      • Local map section or sectors (under the camera).
      • Frustum (sectors visible from the camera).
  • Outputs from Houdini to Dunia:
    • Entity point cloud (vegetation, rocks, collectibles, decals, VFX, etc.).
    • Terrain textures.
    • Terrain height maps.
    • 2D terrain data (RGB or grayscale).
    • Procedurally generated geometry.
    • Terrain logic zones (IDs for post-processing and fog presets).
  • Data Transfer: Outputs are saved as buffers in a specific format for efficient loading by the editor, avoiding large memory transfers.
  • Tool Interconnectivity:
    • Procedural generation is sequential (freshwater, cliffs, biomes, etc.).
    • Tools export data to influence subsequent tools (e.g., freshwater generates a water mask used by the biome tool).

Cliff Tool: Detailed Breakdown

Previous Tech and Motivation

  • Far Cry 4 and Primal had no dedicated cliff tech, relying on terrain and placed rocks.
  • Worlds were designed to minimize large cliff surfaces due to visual limitations and the cost of manual rock placement.
  • Far Cry 5 featured larger and more prominent cliffs, necessitating a new approach.

Cliff Tool Inputs and Geometry

  • Terrain slope is the primary input.
  • Surfaces below a defined slope threshold are deleted, creating the cliff input geometry.
  • Cliffs serve as a visual cue for impassable terrain.
  • Remeshing is applied to eliminate stretched quads on slopes, resulting in uniform triangles.

Stratification

  • Geological stratification (horizontal lines in sedimentary rock) is simulated.
  • Tool slices the input geometry into strata chunks with random thickness.
  • Each strata receives a unique ID for color variation (debug view).
  • Strata angle is controlled by user-painted RGB input on the terrain, with four presets available in the editor.
  • Noise is used to split the cliff surface into two groups, each stratified with a different seed to break up perfect lines and create more natural patterns.

Geometry Generation and Export

  • Strata are extruded with varying thickness and displaced using displacement maps.
  • Triangle count reduction is applied for optimization.
  • Geometry is split into individual meshes per sector (64x64 meters) for improved loading and streaming.

Terrain Shading and Data Transfer

  • Cliffs are shaded using the terrain shader, inheriting textures from the terrain below.
  • To avoid texture mismatch, cliff textures are transferred to the terrain beneath the extruded cliff mesh.
  • Cliff mask and strata attributes are raycast onto the terrain.
  • Strata attribute generates a color layer for macro tint variation, even when cliff mesh is unloaded at a distance.
  • Cliff mask is extended using a flow simulation to create an erosion effect, retaining the original strata color.
  • Crumbled rock entities are scattered on the erosion area and exported as a point cloud.
  • Terrain texture IDs are generated from the erosion mask, using noise to blend two cliff textures for tiling variation.

Vegetation on Cliff Ledges

  • Upward-facing polygons on cliffs are identified as potential vegetation areas.
  • Raycasting checks for clearance above the surface.
  • Trees and other vegetation are spawned on viable cliff ledges.

Cliff Tool Exported Data

  • Cliff geometry with collision information.
  • Point cloud for rocks and ledge vegetation.
  • Terrain texture IDs.
  • Cliff color layer for the terrain.
  • Cliff mask.

Biome Tool: Populating the World with Life

Initial Steps

  • Terrain is generated from the height map.
  • Abiotic data (physical features) is calculated:
    • Occlusion.
    • Flow map.
    • Slope.
    • Curvature.
    • Illumination.
    • Altitude, latitude, longitude.
    • Wind vector map.
  • These attributes form the basis for biome recipes.
  • 2D data from Dunia is imported:
    • Biome painting (user input).
    • Procedural data from previous tools (freshwater, roads, fences, power lines, cliff masks).

Main Biomes and Sub-Biomes

  • Main biomes cover most of the world (75-85%).
  • They automatically determine sub-biome distribution (e.g., forest vs. grassland) based on abiotic terrain data, creating natural macro-level patterns.
  • Main biomes also handle specific rules:
    • Replacing forest with grassland where power lines are present.

Sub-Biome Recipes

  • Sub-biomes are processed sequentially.
  • Each recipe contains ingredients (trees, saplings, bushes, grass).
  • Core of each recipe is the generate-terrain-entities HDA (Houdini Digital Asset), responsible for:
    • Vegetation scattering.
    • Terrain attribute modification.
    • Defining viability for each species.

Viability and Species Competition

  • Viability determines the likelihood of a species growing at a specific location, based on favored terrain attributes.
  • Species with the highest accumulated viability at a location “wins.”
  • Example:
    • Species A favors a specific occlusion range (power of 1).
    • Species B favors a specific flow map range (power of 2).
    • Species B will dominate where the flow map value is sufficiently high.
  • Viability radius determines the influence area of a species, preventing other species from growing too close.
  • Priority and priority radius allow finer control over species interaction:
    • Priority is evaluated first.
    • If priorities are equal, viability is considered.
    • Example: Trees (priority 10) can have a smaller priority radius, allowing bushes (priority 0) to grow closer while still preventing overlap with other trees.

Combining Terrain Attributes for Natural Patterns

  • Mixing various terrain attributes allows for complex vegetation patterns and fluctuating viability.
  • Example: Combining occlusion, altitude, flow map, and noise to create specific growth conditions.
  • Exclusion masks from previous tools (water, cliffs, roads) can be incorporated.

Controlling Asset Size

  • Asset size is linked to viability, allowing for natural size variation within a species.
  • Example: A conifer species can have multiple size variations (50m, 40m, 30m, etc.), each assigned to a specific viability range on the terrain.
  • This creates a tapering effect at forest edges, with smaller trees at the periphery.
  • Scaling percentage for each size prevents a staircase effect, allowing smooth transitions between sizes.
  • Random scale can be added for further variation.
  • Each size can have multiple variations (e.g., a dead tree variant), randomly selected with controllable weights.

Forest Canopy and Ecological Succession

  • Age parameter (signed distance field from viability data) influences size distribution, mimicking forest canopy and ecological succession.
  • It can be added, multiplied, or interpolated with viability to control its influence.
  • Age maximum distance controls the depth of the forest border tapering effect.
  • Age ramp can be used to profile the overall forest shape.

Scattering Density

  • Density ramp controls the scattering density based on size, age, or viability.
  • It prevents overlap issues with larger assets and manages performance by controlling asset numbers.
  • Terrain attributes (illumination, slope aspect) can be mixed into the density ramp for finer control (e.g., reduced density on poorly lit slopes).

Instance Tinting

  • Biomes with high color variation are achieved by tinting individual instances.
  • Gradient color ramps are driven by viability, age, or terrain data.
  • Example: Water signed distance field drives grass color variation near water bodies.

Instance Rotation

  • Asset rotation is primarily based on terrain slope, aligning the forward axis with the slope.
  • This allows for effects like grass leaning towards water due to the shore’s slope.
  • Other options include:
    • Flat horizontal alignment.
    • Wind vector map alignment (e.g., for wheatgrass).
    • Rotation jitter and offset on all axes for random variation.

Terrain Modification by Scattered Assets

  • Scattered assets can influence terrain properties.
  • Four types of terrain data can be modified:
    • Terrain deformation: Height map adjustment (e.g., raising terrain around tree trunks).
    • Terrain textures: Applying specific textures underneath assets (e.g., tree roots).
    • Terrain data output (mask): Generating masks for reuse by other species (e.g., spawning rocks around specific trees).
    • Terrain color: Blending colors with terrain textures (e.g., simulating terrain humidity).
  • Each data type has independent signed distance field settings for controlling the influence area.

Terrain Deformation

  • Masks generated from scattered assets can be combined with terrain data to control deformation.
  • Example: Preventing terrain deformation near roads using the road mask.
  • Displacement height deforms the terrain (e.g., raising it by 1 meter around trees).

Terrain Textures

  • Masks generated from scattered assets determine terrain texture application.
  • Example: Applying a “root” texture underneath trees.
  • Texture selection is dynamically fetched from the engine via Python script.
  • Mask scale controls the area of influence for each texture.

Terrain Data Output and Inter-Species Dependency

  • Terrain data attributes (masks) can be generated from scattered assets and reused by other species.
  • Example:
    • Ponderosa tree outputs a viability mask.
    • Forest rock species uses the Ponderosa viability mask to drive its spawning behavior, creating a dependency between the two.

Terrain Color and Humidity Simulation

  • Terrain tint is generated and mixed with organic terrain textures, similar to the cliff strata color.
  • This creates color variations (e.g., dry brown to lush green) without increasing the number of terrain textures.
  • Neutral gray color allows for both darkening and lightening of the terrain texture.
  • Terrain color and texture variations can transpire through grass, controlled by a mask on the grass asset.

Biome Tool Exported Data

  • Entity point clouds.
  • Terrain texture IDs.
  • Terrain height map.
  • Terrain color.
  • Forest mask (used by fog and wall map tools).

Lessons Learned

  • Responsibility:
    • Procedural tools offer immense control over performance, gameplay, and art, requiring careful consideration of their impact.
    • Balancing art direction (e.g., dense forests) with gameplay needs (e.g., AI navigation) is crucial.
  • Elegant Design:
    • Tools should open up possibilities and be adaptable for various scenarios.
    • The biome tool’s flexibility is highlighted as an example.
  • Simplicity:
    • Avoid over-engineering.
    • Strive for simple and elegant designs through iterative development and refinement.
  • User Feedback:
    • Listen to user needs and preferences.
    • Sometimes manual control is preferred over automation (e.g., riverbed carving).
  • Flexibility:
    • Adapt to changing requirements and be prepared to deviate from initial plans.
  • Balance:
    • Find the right balance between control and automation.
    • Excessive automation can lead to issues, while excessive manual control can be time-consuming and difficult to manage.

Conclusion

  • The procedural pipeline significantly enhanced Far Cry 5’s development, providing efficiency, control, and natural-looking environments.
  • Houdini played a crucial role in achieving the project’s goals.
  • Collaboration and iteration were key to the pipeline’s success.
  • The modular and flexible design allows for future adaptation and reuse in other projects.
About Me:
  • I’m Christian Mills, a deep learning consultant specializing in computer vision and practical AI implementations.
  • I help clients leverage cutting-edge AI technologies to solve real-world problems.
  • Learn more about me or reach out via email at [email protected] to discuss your project.