Model synthesis

Model synthesis a.k.a. wave function collapse or 'wfc' are names for a family of constraint-solving algorithms commonly used in procedural generation, especially in the video game industry.

Some video games known to have utilized variants of the algorithm include Bad North, Townscaper, and Caves of Qud.

The first example of this type of algorithm was described by Paul Merrell, who termed it 'model synthesis' first in his 2007 i3D paper [1] and also presented at the 2008 SIGGRAPH conference and his 2009 PhD thesis.[2] The name 'wave function collapse' later became the popular name for a variant of that algorithm, after an implementation by Maxim Gumin was published in 2016 on a GitHub repository with that name.[3] Gumin's implementation significantly popularised this style of algorithm, with it becoming widely adopted and adapted by technical artists and game developers over the following years.[3]

There were a number of inspirations to Gumin's implementation, including Merrell's PhD dissertation, ideas from quantum mechanics, and convolutional neural network style transfer.[4][5] The popular name for the algorithm, 'wave function collapse', is from an analogy drawn between the algorithm's method and the concept of superposition and observation in quantum mechanics.[6][7] Some innovations present in Gumin's implementation included the usage of overlapping patterns, allowing a single image to be used as an input to the algorithm.[8]

Some have speculated that the reason Gumin's implementation proved more popular than Merrell's, may have been due to the 'model synthesis' implementation's lower accessibility, its 3D focus, or perhaps the general public's computing constraints at the time.[9]

One of the differences between Merrell & Gumin's implementation and 'wave function collapse' lies in the decision of which cell to 'collapse' next. Merrell's implementation uses a scanline approach, whereas Gumin's always selects as next cell the one with the lowest number of possible outcomes.[10]

  1. ^ Merrell, Paul (April 2007). "Example-based model synthesis". Proceedings of the 2007 symposium on Interactive 3D graphics and games (PDF). pp. 105–112. doi:10.1145/1230100.1230119. ISBN 978-1-59593-628-8.
  2. ^ Merrell, Paul (2009). Model Synthesis (PDF). Chapel Hill.{{cite book}}: CS1 maint: location missing publisher (link)
  3. ^ a b Alaka, Shaad; Bidarra, Rafael (2023). "Hierarchical Semantic Wave Function Collapse". Proceedings of the 18th International Conference on the Foundations of Digital Games. p. 2. doi:10.1145/3582437.3587209. ISBN 978-1-4503-9855-8. In 2016, Maxim Gumin unleashed the WFC algorithm, publishing a repository containing his initial implementation. Since then, WFC has had a profound impact on technical artists and game developers, getting adopted, adapted and used in commercially published and upcoming projects (Caves of Qud, Townscaper, Matrix Awakens).
  4. ^ Merrell, Paul (Aug 6, 2023). Procedural Modeling Using Graph Grammars (Video). Event occurs at 3:13.
  5. ^ "Implementing Wave Function Collapse & Binary Space Partitioning for Procedural Dungeon Generation". Shaan Khan. 2021-03-21. Retrieved 2024-03-24. In the case of WFC it is inspired off three distinctive but functionally similar algorithms& concepts; Texture Synthesis (Specifically Discrete Synthesis), Markov Chains & Quantum Mechanics. WFC was also additionally inspired by convolution neural network style transfer (CNN Style Transfer).
  6. ^ Gumin, Maxim (September 2016), Wave Function Collapse Algorithm, retrieved 2024-03-24
  7. ^ "The Wavefunction Collapse Algorithm explained very clearly". Robert Heaton. Retrieved 2024-03-24.
  8. ^ Gumin, Maxim (September 2016), Wave Function Collapse Algorithm, retrieved 2024-03-25
  9. ^ Alaka, Shaad (2023). "Hierarchical Semantic Wave Function Collapse". Proceedings of the 18th International Conference on the Foundations of Digital Games. p. 2. doi:10.1145/3582437.3587209. ISBN 978-1-4503-9855-8. Years before, Merrel had published the conceptually identical Model Synthesis algorithm, though it did not catch on as much as WFC did, possibly due to its lower accessibility, main 3D focus and computing requirements at the time.
  10. ^ Merrell, Paul (28 July 2021). "Comparing Model Synthesis and Wave Function Collapse" (PDF). The first difference is in the step where we choose a cell and pick a label. The cells are chosen in a different order. Model synthesis sweeps through the grid in scanline order. WFC chooses the lowest entropy cell.