seleCte

Generates Celeste levels procedurally that the game can load and you can eventually play.

seleCte learns the tile patterns of Celeste's 824 vanilla rooms with a multi-dimensional Markov chain, draws a level layout, fills each room by sampling from the model, and encodes the result as a .bin map. It started as a Master's thesis and is an active personal side-project: the levels load, start on a player spawn and end on a crystal heart, but they are not yet reliably good. See Limitations.

Quickstart

You need uv, which also installs the right Python for you. You do not need a copy of Celeste: the trained model ships with the repository.

git clone https://github.com/robilouis/seleCte
cd seleCte
uv sync
uv run python -m seleCte.pcg.level_generation --nb-rooms 7 --proba 0.3 --room-size 22 14 --tries-limit 4 --level-name my_level

After a few seconds you have my_level/: one room_<n>.csv per room, and a data.json describing the layout.

Play it

Turning the level into a map needs Julia 1.11 or later:

julia --project=. -e "using Pkg; Pkg.instantiate()"
julia --project=. src/seleCte/room_encoder/data_to_level.jl my_level my_level

That writes generated_bin_files/my_level.bin. Drop it into an Everest mod folder and pick it from the chapter list.

Options worth knowing

Flag	What it does
--nb-rooms	How many rooms the level has.
--proba	Layout shape: 0 is a straight corridor, values near 1 a labyrinth.
--room-size	Width then height of each room. The default 40 23 works but takes minutes.
--level-name	Output folder.
--reset-skeleton	Redraw the whole level when a room never comes out playable.

Every script prints its full options with --help.

How it works

your copy of Celeste ──► extract levels ──► train the Markov model ──► model (committed)
                                                                              │
                     playable .bin ◄── encode ◄── fill rooms ◄── draw layout ◄┘

The right-hand half is all you need to generate levels. The left-hand half is only for retraining, and needs the game. docs/how-it-works.md walks through each stage, the modules, and the tile alphabet.

Limitations

• "Playable" is approximate. The check looks for a path through empty space and ignores gravity, jumping, dashing and stamina. It accepts rooms that cannot be played and rejects some that can. Replacing it is the main line of active work.
• Large rooms are slow. The playability check scales badly with room size.
• Nothing that moves. Zip movers, moving blocks, falling blocks and the like are dropped.
• Partial entity coverage. About a tenth of Celeste's entities are supported; lightning and springs are the notable gaps.
• Backgrounds stop short. Background tiles can be generated but not yet written into a map.

Known issues and planned work are in the issue tracker.

Retraining the model

This needs your own copy of Celeste. No game data is shipped here.

julia --project=. src/seleCte/data_loader/data_loader.jl "<Celeste>/Content/Maps" src/seleCte/data/levels/original
uv run python -m seleCte.pcg.preprocessing.recombine_fg_entities
uv run python -m seleCte.pcg.dpe_generation --dataset 0123456789L

The first command extracts every room from the game's maps, the second turns them into training data, and the third trains a new model. Details, including the background layer, are in docs/how-it-works.md.

Development

git add -A
uvx pre-commit run --all-files
uv run mypy --no-pretty
uv run pytest -q
julia --project=. tests/julia_smoke.jl

Stage first: pre-commit only sees files git knows about. All four checks also run in CI.

Citing

This repository is the implementation behind:

Louis Robinet, Marcello A. Gómez-Maureira, and Mike Preuss. 2025. Towards a Celeste AI Framework: Agent-free Automated 2D Level Generation for Multidirectional Platformers. In FDG '25: Proceedings of the 20th International Conference on the Foundations of Digital Games. https://doi.org/10.1145/3723498.3723796

GitHub's Cite this repository button gives the BibTeX entry.

The paper describes an earlier version of the code, from December 2024. It has been substantially reworked since; CHANGELOG.md lists what changed.

License

MIT, see LICENSE.

Celeste is © Maddy Makes Games / Extremely OK Games. This is an unaffiliated, non-commercial research project, and no game assets or level data are redistributed here. The Julia scripts build on Maple by the Celestial Cartographers, pinned to a fork that can encode the crystal heart.