Procedural generation

See also: Text-to-3D generation and Skybox AI

Procedural generation, also called procedural content generation (PCG), is the creation of game content by algorithms instead of by hand. The introductory chapter of the textbook Procedural Content Generation in Games (Springer, 2016) defines PCG as "the algorithmic creation of game content with limited or indirect user input". Content in this definition includes levels, maps, game rules, textures, stories, items, quests, music, weapons, vehicles and characters, but not the game engine or the behavior of non-player characters. Generators range from hand-written rules to search algorithms, constraint solvers and models trained on existing content, including large language models.

Procedural generation is used in virtual reality games such as No Man's Sky, whose generated planets became explorable in VR with the Beyond update in August 2019, and Elite Dangerous, whose galaxy of 400 billion star systems is constructed from real astronomical data. Roguelite VR games such as In Death and Until You Fall randomize their maps or rooms for each run. When VR users walk naturally, the virtual environment has to fit inside the tracked physical space. Research systems such as flexible spaces (2013), a system by Misha Sra and colleagues (2016) and VRoamer (2019) therefore generate virtual layouts that fit a tracked workspace, a scanned room or a building through which the user walks.

Some game engines include procedural tools. Epic Games added an experimental Procedural Content Generation framework to Unreal Engine 5.2 in May 2023 and declared it production-ready in Unreal Engine 5.7 in November 2025. Research presented in 2023 and 2024 used large language models to generate game levels and interactive mixed reality scenes from text prompts.

Reviewed 1 November 2026. Checked every cited paper's DOI record and its abstract or full text, the PCG textbook chapters, the WFC README, Epic's Unreal Engine 5.2 to 5.8 release posts and the developer, store and press sources for the game rows About review dates.

Definition

The authors of the textbook call their definition "somewhat arbitrary and rather fuzzy around the edges" and note that some researchers prefer the term "generative methods" for a wider set of techniques. They stress that generated content must be playable: it should be possible to finish a generated level or to use a generated weapon. A 2013 survey by Mark Hendrikx and colleagues sorts game content into a "six-layered taxonomy of game content: bits, space, systems, scenarios, design, and derived" and finds that "many of the methods used to generate game content from one layer can be used to generate content from another". The survey describes the task as difficult because a generator has to create the content, satisfy constraints imposed by the artist and return results that interest players.

The textbook classifies generation methods along seven dimensions, in a revised version of a taxonomy first presented by Togelius and colleagues. A given method usually lies somewhere between the two ends of each dimension.

Dimension	Meaning	Example given in the textbook
Online or offline	Content is generated while the game is played, or during development or before a session	Left 4 Dead (Valve, 2008) analyzes player behavior on the fly and alters the game state with PCG techniques
Necessary or optional	Content required to complete a level, which should always be correct, or auxiliary content that can be discarded or exchanged	The main structure of Super Mario Bros. levels is necessary content; weapons in first-person shooters are optional content
Degree of control	Generation is steered by a random seed or by parameters along several dimensions	Minecraft regenerates the same world from the same seed
Generic or adaptive	Content is generated without regard to player behavior, or on the basis of it	Galactic Arms Race evolves the weapons it offers from the player's earlier weapon use; according to the authors, most commercial games use generic PCG
Stochastic or deterministic	Whether the same content can be regenerated from the same starting point and parameters	Elite regenerates its galaxies deterministically
Constructive or generate-and-test	Content is produced in one pass, or generated and tested in a loop until a result is satisfactory	Roguelike games commonly generate in one pass; the board game Yavalath was generated by a program using generate-and-test
Automatic or mixed authorship	Generation with limited designer input, or a designer or player working together with the algorithm	In Tanagra, a designer draws part of a 2D level and a constraint satisfaction algorithm generates the missing parts while keeping the level playable

The same chapter lists properties that are often wanted from a generator: speed, with requirements that range from milliseconds to months depending, among other things, on whether content is generated during play or during development; reliability; controllability; expressivity and diversity; and creativity and believability, because in most cases the content should not look as if a procedural content generator designed it. These properties can conflict, for example speed with quality, or expressivity and diversity with reliability. Reliability matters more for some content than for other content: in the authors' example, "a dungeon with no exit or entrance is a catastrophic failure", whereas an odd-looking flower does not necessarily break the game.

History

According to the textbook, overcoming the storage limits of computers was one of the main reasons the techniques were developed. Elite (1984) stored the seed numbers from which it generated eight galaxies of 256 planets each, and the dungeon game Rogue generated new levels every time a game started. The roguelike genre follows Rogue (1980) in generating its dungeons at runtime. Commercial games that the textbook names as users of PCG include Diablo, which generates its maps and the type, number and placement of items and monsters; Spore, which animates player-designed creatures procedurally and places them in a generated galaxy; Civilization IV, with random maps; Minecraft, which generates its whole world and its content; and Spelunky, with generated variations of 2D platform levels. The authors credit Will Wright's talk "The Future of Content" at the 2005 Game Developers Conference with helping to revive interest in procedural content generation.

Ken Perlin's 1985 paper "An image synthesizer" in ACM SIGGRAPH Computer Graphics described building "naturalistic visual complexity" by composing non-linear functions and introduced the concept of "solid texture". In 1997 the Academy of Motion Picture Arts and Sciences gave Perlin a Technical Achievement Award "for the development of Perlin Noise, a technique used to produce natural appearing textures on computer generated surfaces for motion picture visual effects". The PCG textbook calls the diamond-square and Perlin noise methods "classic and widely used" for terrain generation. Writing in 2013, Hendrikx and colleagues described "a large body of research" on procedural content generation for games, "particularly over the past decade".

Methods

The families below follow the cited textbook and surveys; many generators combine several of them.

Family	Principle	Typical content in the cited sources
Constructive	Runs in fixed, usually short time and does not evaluate its output in order to regenerate it	Dungeons, caves and platform levels
Noise and fractals	Fills a grid with smoothly varying random values, often at several scales	Terrain heightmaps, cloud cover in skyboxes
Search-based	Evolutionary or other metaheuristic search, guided by an evaluation of each candidate's quality or fitness	Content for digital games and for board games
Constraint solving	Searches for output that satisfies constraints, such as local similarity to an example or playability	Bitmaps, tile maps and voxel models (WaveFunctionCollapse); 2D levels (Tanagra)
Machine learning (PCGML)	Models trained on existing content generate new content	Platformer levels, game maps, interactive fiction stories, collectible card game cards
Large language models	A fine-tuned or prompted language model generates content, or code that builds it, from text	Super Mario Bros. levels (MarioGPT); interactive Unity scenes (LLMR)

Constructive methods

Chapter 3 of the textbook covers dungeon and level generators that it calls constructive: they "run in fixed (usually short) time, and do not evaluate their output in order to re-generate it". Binary space partitioning, which the chapter calls "the most popular method for space partitioning", recursively divides a space into two subsets, and the chapter applies such partitions to dungeon layouts. Agent-based generators let agents dig paths "into a primeval mass of matter", cellular automata are "a simple and fast means of generating structures such as cave-like dungeons", and generative grammars can capture higher-level aspects of dungeon design. The chapter ends with the platform level generators of the commercial game Spelunky and the open-source Infinite Mario Bros. framework.

Noise, fractals and terrain

Chapter 4 represents noise and most aspects of terrain as two-dimensional grids of real numbers: for terrain, a heightmap stores the height at each point, and for noise the equivalent grid is called an intensity map. Filling a heightmap with independent random values gives what the chapter calls "random spikes, not random terrain". Interpolated (value) noise instead generates random heights on a coarser lattice and interpolates between them, and gradient noise, "sometimes called Perlin noise", interpolates random gradients, which adds a further level of smoothness. Real terrain varies at several scales, and the chapter calls generated terrain with this kind of self-similarity fractal terrain. The diamond-square algorithm, which the chapter describes as commonly used in games because it is relatively simple to implement and computationally efficient, repeatedly subdivides a square and adds random values; their magnitude, called the roughness, is reduced at each step.

Heightmaps cannot represent caves or overhanging cliffs. A voxel grid can, at the cost of much more storage, and the chapter cites Minecraft as a game "which uses unusually large voxels". Noise also has uses outside terrain, for example cloud cover in skyboxes. The chapter notes that, in the forms it presents, noise and agent-based methods "suffer from a certain lack of controllability": it is not easy to require, for instance, that two points on a terrain be easily reachable from each other.

Search-based generation

Julian Togelius, Georgios N. Yannakakis, Kenneth O. Stanley and Cameron Browne proposed the term "search-based procedural content generation" in a 2011 survey of research that applies "evolutionary and other metaheuristic search algorithms" to generating game content. Their taxonomy classifies such generators by what kind of content is generated, how it is represented and how its quality or fitness is evaluated. In a VR example described below, Sra and colleagues used a genetic algorithm to place virtual objects within the walkable area of a scanned room.

Constraint solving and WaveFunctionCollapse

WaveFunctionCollapse (WFC), a program by Maxim Gumin published on GitHub, "generates bitmaps that are locally similar to the input bitmap": the output may contain only NxN patterns of pixels that occur in the example, in a similar distribution. The program repeatedly picks the unobserved region with the lowest Shannon entropy, collapses it to a definite state and propagates the consequences through the output. In its "simple tiled model", which starts from a list of tiles and their adjacency data, propagation "is just adjacency constraint propagation"; the algorithm works the same way in three dimensions, where the repository shows generated voxel models. Generation can run into a contradiction, after which it cannot continue. The README notes that the underlying problem is NP-hard, "so it's impossible to create a fast solution that always finishes", although in practice contradictions occur "surprisingly rarely".

Isaac Karth and Adam M. Smith, at the 2017 Foundations of Digital Games conference, analyzed WFC as "an example-driven image generation algorithm" that implements "a non-backtracking, greedy search method", explained it with ideas from the constraint solving literature and built a surrogate implementation with answer set programming. They date the program's release to 30 September 2016 and describe WFC attracting the attention of several indie and hobby game makers through social media in 2016 and 2017. As of November 2026 the README lists Bad North, Caves of Qud, Dead Static Drive and Townscaper among the games whose levels WFC generates, and links adaptations for Unity, Unreal Engine 5, Godot 4 and Houdini.

Machine learning and language models

A 2018 survey by Adam Summerville and colleagues defines procedural content generation via machine learning (PCGML) as "the generation of game content using machine learning models trained on existing content", in contrast with search-based, solver-based and constructive methods. It concentrates on functional content such as platformer levels, game maps, interactive fiction stories and cards in collectible card games, and notes that modeling existing content also suits repair, critique and content analysis. The methods it covers include neural networks (among them long short-term memory networks, autoencoders and deep convolutional networks), Markov models such as n-grams and multi-dimensional Markov chains, clustering and matrix factorization. Open problems listed in the survey include learning from small data sets, lack of training data, style transfer, parameter tuning and PCG as a game mechanic.

MarioGPT, presented at NeurIPS 2023 by Shyam Sudhakaran and colleagues, fine-tunes a GPT-2 model to generate tile-based Super Mario Bros. levels. According to the authors, it "can be text-prompted for controllable level generation", and combined with novelty search it generates diverse levels with varying player paths. Generative models for 3D assets and scenes, and for 360-degree panoramas, are covered in Text-to-3D generation and Skybox AI.

Engine and tool support

Epic Games announced its in-house Procedural Content Generation framework (PCG) as an experimental feature of Unreal Engine 5.2. According to Epic, the framework offers the ability to "populate large scenes with Unreal Engine assets" and includes both in-editor tools and a runtime component. Selected later releases extended it:

Release	Announced	PCG framework changes, as described by Epic
Unreal Engine 5.2	11 May 2023	Introduced as an experimental feature, with in-editor tools and a runtime component
Unreal Engine 5.6	3 June 2025	Better GPU-driven performance and GPU-based spawning at runtime, multithreading, a 3D viewport for previewing points, textures and meshes, and the PCG Biome Core v2 plugin with per-biome blending and biome layering
Unreal Engine 5.7	12 November 2025	Declared production-ready; a new PCG Editor Mode with tools to draw splines, paint points or create volumes, each linked to a PCG graph; faster GPU compute; and the first experimental release of the Procedural Vegetation Editor, which outputs Nanite skeletal assemblies
Unreal Engine 5.8	17 June 2026	Updates for nondestructive manual editing, richer procedural environments and streamlined graph workflows; the Procedural Vegetation Editor creates Nanite-ready vegetation assets that react to wind systems

SideFX's Houdini Engine loads Houdini Digital Assets, Houdini node networks encapsulated into single nodes, into 3D applications such as Autodesk Maya and 3ds Max and game editors such as Unity and Unreal Engine. Parameters added to an asset can be edited in the host application, and changes to the core asset update every instance of it. As of November 2026, SideFX offers Houdini Engine for Unreal and Unity free to commercial customers, with up to 10 licenses per studio through its website, and a free Indie version for up to three computers.

Use in VR and AR games

The table lists VR games, and one fan-made tool, whose procedural or generated content is documented by the developer or by the press. It is not a complete list.

Game	Developer	VR support	Generated content, as described by the source
No Man's Sky	Hello Games	Beyond update, released 14 August 2019, for PC VR headsets and PlayStation VR; PlayStation VR2 support from the Fractal update (version 4.1) on 22 February 2023	Planets; Road to VR cites over "18 quintillion" procedurally generated planets
Elite Dangerous	Frontier Developments	VR supported on Steam; announced in March 2016 as an Oculus Rift launch title	A Milky Way of 400 billion star systems constructed from real astronomical data
Minecraft (Bedrock Edition)	Mojang	Oculus Rift support from 2016 and Windows VR headsets from 2017; Bedrock Edition patch notes confirmed the end of VR and mixed reality support in May 2025	The whole world, which can be regenerated from the same seed
Microsoft Flight Simulator (2020)	Asobo Studio	VR through OpenXR from December 2020	1.5 billion buildings reconstructed from 2D satellite images for areas outside the 400 cities covered by Bing Maps photogrammetry
In Death	Sólfar Studios	VR only; Steam release date 2 October 2018	A map generated from a random seed for each run, with randomized enemy and loot spawn points
Until You Fall	Schell Games	VR only; Steam release date 27 October 2020	Randomized rewards, power-ups, enemy configurations and rooms in each run
COMPOUND	Bevan McKechnie	SteamVR; released on 19 July 2022 after Steam Early Access	Randomized level layouts
Ancient Dungeon	Eric Thullen	VR only; Steam release date 3 November 2021, listed in the Early Access genre	Dungeons
Dungeons of Eternity	Othergate	VR only; Steam release date 11 September 2025	Randomly generated dungeons across several game modes and realms; randomly generated weapons
Audioshield	Dylan Fitterer	HTC Vive, released on the headset's launch day, 5 April 2016	Game levels converted from the player's music files
Beat Sage, a fan-made map generator for Beat Saber	Chris Donahue and Abhay Agarwal	Custom maps for Beat Saber on SteamVR headsets and Oculus Quest, where installation requires modding through SideQuest	Beat maps generated by two neural networks from a song on YouTube or an mp3 file

At the 2017 Game Developers Conference, Innes McKendrick of Hello Games presented "Continuous World Generation in 'No Man's Sky'". The session description says that the game "uses procedural generation by necessity" to create game worlds in which a player can move from space down to an interactive and populated terrain, with planets generated continuously in real time, and that the talk breaks down a pipeline "from voxel-based world generation, through polygonization and texturing, to eventual population and simulation". Hello Games states that the Beyond update made the whole game playable in VR, with Oculus Touch, Valve Index or Vive controllers on PC, and a DualShock 4 or PlayStation Move controllers on PlayStation VR. Hello Games released the Fractal update (version 4.1) on 22 February 2023 to coincide with the launch of PlayStation VR2, with full support for that headset.

In Elite Dangerous, everything in the game's Milky Way is generated "according to real science" by a system called the Stellar Forge, PC Gamer reported in 2018. Kay Ross, a senior programmer at Frontier, told the magazine: "Before a planet can be made, you must first simulate the circumstances of its birth." Planets begin as a disc of material around a newly formed star. The Hipparcos and Gliese star catalogues supplied information about observable stars, and missing information is filled in by calculation. The engine generates topological features "from high orbit down to human scale", while artists add material detail and can insert custom topography for gameplay, story or settlements. In March 2016 PC Gamer reported that the game would be released for the Oculus Rift as a launch title on 28 March.

For Microsoft Flight Simulator (2020), developed by Asobo Studio, Bing Maps photogrammetry data yielded exact replicas of only 400 cities, according to TechCrunch. For the rest of the planet, the Graz-based startup Blackshark.ai, a spin-off of the game studio Bongfish, reconstructed 1.5 billion buildings from 2D satellite images with a team of about 50 people. Its chief executive, Michael Putz, described finding building footprints as a computer vision task and said that machine learning completes buildings partly hidden by the shadows of trees. A free update added VR through OpenXR in December 2020, with support for most Windows Mixed Reality headsets, including the HP Reverb G2, and for Oculus, Valve and HTC headsets.

Beat Sage's creators state that its neural networks were "trained on Beat Saber levels created by humans": the first predicts at what points in time blocks should be placed, and the second assigns a block type to each predicted time. In augmented reality, the mobile game Peridot generated each of its digital pets ("Dots") through a system called Perigenetics: according to the game's launch announcement, "each Dot is procedurally generated with 100% unique DNA", and the number of possible variations of creature looks is 2.3 x 10²⁴. In April 2026 Niantic Spatial announced that it was winding down the mobile game: the app was to be removed from the App Store and Google Play on 14 May 2026, and the servers were to close on 31 August 2026. The announcement says that the franchise continues in the headset mixed reality experience Hello, Dot and in Peridot Beyond on Snap Spectacles.

Generating environments that fit physical space

In room-scale VR and other setups where users walk, Evan A. Suma and colleagues summarized the basic limit in 2012: "Walking is only possible within immersive virtual environments that fit inside the boundaries of the user's physical workspace." Redirected walking techniques usually manipulate "the mapping between physical and virtual motions while the layout of the environment remains constant"; the systems below instead design, generate or change the virtual layout itself.

System	Publication	Physical input	What is generated or changed
Impossible spaces	IEEE Transactions on Visualization and Computer Graphics, 2012	A tracked workspace of 9.14 m x 9.14 m	Self-overlapping architectural layouts that compress large interiors into a smaller physical area
Flexible spaces	IEEE Symposium on 3D User Interfaces, 2013	A tracked workspace	Procedurally generated layouts that combine overlapping spaces with change blindness, for unlimited real walking
Sra and colleagues	ACM VRST, 2016	A 3D map of an indoor space made with a Project Tango tablet	A virtual world fitted to the walkable area, with detected real objects such as a chair replaced by virtual counterparts
VRoamer	IEEE VR, 2019	Depth camera data captured while walking through a building, without prior scanning	Pre-authored virtual rooms where they fit, otherwise virtual corridors and doors leading to undiscovered areas

Impossible spaces were introduced by Suma, then at the USC Institute for Creative Technologies, with Zachary Lipps, Samantha Finkelstein, David M. Krum and Mark Bolas. In their first experiment, small virtual rooms could overlap by as much as 56% before users began to detect that they were in an impossible space, and larger rooms that filled the 9.14 m x 9.14 m workspace could overlap by up to 31%. Users judged distances to objects in adjacent overlapping rooms as if the space were not compressed, even when the overlap was noticeable. A second experiment chained impossible spaces in an expansive outdoor scene with several established redirection techniques, and users' comments suggested that the illusion was stronger when they did not know about the manipulation. Flexible spaces (2013), by Khrystyna Vasylevska, Hannes Kaufmann, Bolas and Suma, generalizes overlapping spaces and change blindness "by employing procedural layout generation", so that designers can "focus on the content of the virtual world independent of the implementation details imposed by real walking".

Sra and colleagues, at the MIT Media Lab and the University of Córdoba, generated VR worlds from a 3D reconstruction of the user's surroundings. A Project Tango tablet served both as the 3D mapping device and as the head-mounted display. The system detects obstacles such as furniture and walls and the walkable area in the point cloud, then a genetic algorithm (population of 1,000, mutation probability 0.05, at most 1,000 iterations) places virtual elements. A detected real chair can be replaced by a virtual one that the user touches and sits on; unlike earlier passive haptics approaches, in which physical objects were built and placed in the room, the system detects objects that are already there. Processing a point cloud and generating a world took about 2 to 4 minutes per space, including the initial mapping. The object recognition pre-process took 20 minutes to an hour, depending on the object, and the system ran at 25 to 30 frames per second with one tracked object. Some users ignored the virtual boundary elements and had to be steered away from physical obstacles; in more than 20 tests no subject reported nausea.

VRoamer, presented at IEEE VR 2019 by Lung-Pan Cheng and colleagues at Microsoft Research, needs no scan before use. A backpack PC with an NVIDIA GTX 1080 drives an Acer Windows Mixed Reality headset with inside-out tracking, and an Intel RealSense D435 depth camera with a range of 30 cm to 10 m detects walkable areas and obstacles in real time. The walkable area is eroded by 50 cm, which the authors describe as a safety margin. VRoamer instantiates pre-authored virtual rooms where they fit the walkable area and otherwise generates corridors and doors that lead to undiscovered parts of the building; the authors contrast this with "redirected walking techniques that fold virtual spaces". In a study, 14 participants followed a virtual fairy through a generated dungeon while avoiding virtual obstacles. Their mean walking speed was 0.74 m/s, over a mean distance of 66.23 m. Ten participants stopped before entering a narrow virtual corridor where the physical path became narrow, and one collided with an obstacle while side-stepping. The authors report a high mean presence score of 5.5 out of 7, and none of the participants correctly guessed their final location in the building.

Other VR research

VREX, a free open-source Unity toolbox for virtual reality research in experimental psychology and neuroscience, published in BMC Psychology in 2017, "provides a procedural generation of (interconnected) rooms that can be automatically furnished with a click of a button"; researchers can also place and adjust objects by hand in its editor.

ProGenVR, from the University of Porto and INESC TEC (ICAT-EGVE 2022), brings rule-based PCG into a VR modeling tool. A user builds a main model, such as a desk or a city block, from an asset library with a set of procedural rules and clones it; every clone follows the same rules but differs through randomization. The prototype was developed in Unity with SteamVR and an Oculus Rift headset. In an evaluation with 15 participants (one female, aged 16 to 30), who modeled a computer room and a town center, the authors compared it with a baseline restricted to a core subset of objects and operations. According to the authors, the results suggest that the time and mental effort of defining rules can be compensated by the time and physical effort saved when creating complex scenes.

Tobias Huber and colleagues at the University of Augsburg (IEEE SSCI 2021) combined procedural generation with dynamic difficulty adjustment in a VR exergame with exercises intended to prevent lower back pain. The mazes, which contain exercise rooms, are generated by a neural network trained with deep reinforcement learning to adapt their difficulty to the player; the prototype used an HTC Vive Pro. The authors describe the results of their exploratory study with 19 students as promising but to be taken only as a proof of concept, and they report conflicting results for participants who were already satisfied with the first maze. Participants spent up to 10 minutes in each maze, which the authors say drastically reduced the number of learning steps possible in a given time and slowed adaptation. They also state that more participants would be needed to statistically verify the usability of a finished VR exergame that uses the approach.

LLMR (Large Language Model for Mixed Reality), presented at CHI 2024 by Fernanda De La Torre, Cathy Mengying Fang, Han Huang, Andrzej Banburski-Fahey, Judith Amores Fernandez and Jaron Lanier, all affiliated with Microsoft, creates and modifies interactive mixed reality scenes from prompts. A Planner splits a request into a sequence of sub-prompts, a Scene Analyzer summarizes the current scene, a Skill Library helps guide a Builder that writes Unity C# code, and an Inspector checks that code for compilation and run-time errors before the Roslyn compiler compiles it and Unity runs it. Plug-ins extend the system, for example to load models from Sketchfab. In the evaluation, LLMR is GPT-4 augmented with these modules. On two sets of 150 prompts, each run five times, its errors fell to 20.5% and 25.2% of plain GPT-4's error rate for an empty scene and an existing bathroom scene, which the authors summarize as a fourfold improvement in average error rate. It took longer: 90.98 seconds on average for a single prompt in the empty scene against 35.24 seconds for GPT-4, on a PC with 32 GB of RAM and an RTX 3080 GPU. The 11 participants of a usability study found it intuitive and easy to use. The authors deployed it on devices such as Meta Quest and Microsoft HoloLens 2; to keep the framework lightweight, its run-time compiler runs on a PC that acts as a server, and the generated results are streamed to the device.

Limitations

Repetition is a documented problem. Younès Rabii and Michael Cook (FDG 2023) discuss the Oatmeal problem, which they describe as "a term coined by Kate Compton": "Every bowl of oatmeal is unique, Compton explains, but that does not make them interesting or valuable." Designers use the analogy to understand that "procedural generation alone does not guarantee variety or interest". The paper proves a relationship between the Kolmogorov complexity of the most complex artifact a generator can produce and the size of its possibility space. It also notes that players learn to identify patterns and complain about reused assets in games generally, but that procedurally generated content is "more susceptible to pattern identification". In a VR example, UploadVR's 2018 review of In Death found that each run "really does feel like a different experience each time", although "the walls, floor, objects in the environment, and textures all mostly look about the same", and named recycled assets among the game's issues.

Control and evaluation are further problems. Noise-based terrain methods are hard to control, as described above. To evaluate generators, Gillian Smith and Jim Whitehead described a method in 2010 for analyzing the expressive range of a level generator, which examines the variety of levels it produces and the effect of changing its input parameters. Applying it to Launchpad, a level generator for 2D platformers, they found that such analysis "can also expose unexpected biases in the generation algorithm and holes in the expressive range". Constraint solvers such as WFC can stop at a contradiction. At large scale, manual checking becomes impractical: Putz told TechCrunch that with 1.5 billion buildings "you cannot do traditional Q&A anymore" and that "it might be that 20% of the buildings are off", which he guessed was the case in the flight simulator. Learned generators reflect their training data: UploadVR reported that the sample tracks behind Beat Sage's model mainly feature fast electronic music, so pop and electronic songs often give better maps, while results for rock music and for songs with a slow tempo or sparse instrumentation are mixed.

Systems that fit generated worlds to physical space depend on sensing. The system by Sra and colleagues was limited to spaces with a constant floor height, without stairs or ramps; tracking could drift with fast movements, in dimly lit spaces or in spaces without features; and it did not track other people or pets. The VRoamer authors note that a longer sensor range reduces the need to slow users down, that depth cameras may not sense transparent and reflective surfaces accurately, and that a room the user returns to is regenerated at a new location if its original physical space is no longer available. For LLM-based generation, the LLMR authors list unpredictability "due to generative models' stochastic nature" and state that the system is not suited to contexts that require precise and specific control; its Inspector module nearly doubles code generation time.

See also

• How to develop VR games with AI
• Text-to-3D generation
• Skybox AI
• Dynamic difficulty adjustment
• Redirected walking
• Level of detail
• Game engine