Generating more is easy to demonstrate.
The study asks whether a system can produce useful variation at scale without obvious repetition, broken composition, or unmanageable graphs.
Applications are open for the January 2027 cohortApply
AI & Procedural Systems
Studying whether intelligent and rule-based production tools can expand a world without erasing the decisions that make it authored.
The central research question
Can an AI-assisted procedural pipeline build a production-ready environment faster without weakening art direction, traceability, or runtime performance?
Research scope
Procedural generation is already capable of placing vegetation, buildings, roads, and entire biomes. New language-driven interfaces can also help people construct and modify procedural graphs. The unresolved question is not whether these systems can generate more content. It is whether the complete production process becomes better after correction, review, optimization, and artistic revision are counted.
Inviox will study directed generation: systems constrained by a human-authored brief, asset library, spatial grammar, performance budget, and approval process. The research will compare the total cost and quality of a procedural workflow with a conventional hand-authored workflow under the same conditions.
Why this question
Generation is not the same as production.
A useful system must survive the parts of development that demonstrations usually leave out.
Artists must be able to direct and revise the result.
Every generated decision should remain inspectable, editable, reproducible, and attributable to a rule, input, or approved tool.
Fast output can create slow cleanup.
Setup, failed generations, manual correction, technical review, optimization, and later art-direction changes will all be included in the production cost.
Initial Inviox study
The Directed Worldbuilding Trial
Two equivalent real-time districts will be produced from the same brief and asset library: one through a conventional hand-authored workflow and one through a directed procedural workflow. A crossover design will reduce the effect of individual artist speed.
How the study will run
- 01
Define the claim
Preregister the research question, success criteria, failure conditions, and the evidence required before building begins.
- 02
Lock the production brief
Use the same 150 × 150 metre district, asset kit, reference board, hardware, visual target, and frame-time budget for both workflows.
- 03
Build the procedural grammar
Create deterministic rules for roads, building types, density, sightlines, prop placement, variation, exclusions, and artist overrides.
- 04
Run a crossover production test
The same practitioners complete both workflows in a balanced order while time, revisions, failures, and interventions are logged.
- 05
Review without revealing the method
Independent art and technical reviewers score composition, coherence, repetition, editability, and runtime quality without knowing which workflow produced each result.
- 06
Publish the complete cost
Report useful gains, cleanup time, failed approaches, limitations, and the conditions under which procedural assistance should not be used.
What will be measured
- Total practitioner hours
- Time required for a major art-direction revision
- Blind art-direction score
- Visible repetition and rule violations
- Manual correction load
- Frame time, memory, draw calls, and build size
- Percentage of outputs with complete provenance
Research controls
- Identical brief and approved asset library
- Identical target hardware and performance budget
- No confidential or unlicensed training material
- Human approval for every production-facing result
- Versioned graphs, prompts, seeds, overrides, and review notes
Research value
Planned outputs
This page does not claim a result. It defines the evidence and public outputs the study is expected to produce.
Working prototype
A reproducible Unreal Engine environment and procedural graph demonstrating the tested workflow.
Production benchmark
A transparent comparison of time, revision cost, quality, performance, and provenance.
Decision guide
A practical guide explaining where procedural assistance helps, where it creates debt, and where human authorship must remain direct.
Teaching case
A documented brief that students can inspect, challenge, and repeat rather than a promotional technology demonstration.
Evidence base
Existing work that frames the question
These sources demonstrate current capabilities. They do not answer Inviox’s production comparison for us.
INVIOX / RESEARCH
Contribute a production question.
Studios and practitioners can propose a workflow, review criterion, or production constraint for the study.