Mario Maker Procedural Generation

Setting up

Run pip install -r requirements.txt

The conversion pipeline (mm2pipeline_data)

Every format conversion (HuggingFace dump -> .bcd -> JSON/PNG -> ASCII -> dataset -> generated ASCII -> JSON -> playable .swe) lives in the mm2pipeline_data package and runs through one entry point:

python -m mm2pipeline_data <command>     (extract | toost | json-to-ascii | dataset | ascii-to-json | swe)

See mm2pipeline_data/README.md for full instructions and walkthroughs. Common one-liners:

Build a dataset from ASCII level files: python -m mm2pipeline_data dataset build --input <input.txt or folder> --output_folder <dataset_name>.json --tileset extended_tiles.json --sliding_window --stride 20 --convert_to_extended

Convert a level .json to SWE (playable in SMM: World Engine): python -m mm2pipeline_data swe --input -o <output .swe path or folder> --user

Running Bat Files

Extract real MM2 levels to ASCII (bcd -> json -> images -> ascii) bat\extract_levels_to_ascii.bat <name_filter> <output_folder>

Prepare a Mario Maker dataset end-to-end (simple presence captions): bat\prepare-mario-maker.bat <folder of ascii .txt>

Prepare a Mario Maker dataset end-to-end (LLM captions via local Ollama): bat\prepare-mario-maker-llm.bat <folder of ascii .txt>

Run the full pipeline (prepare + train MLM + train diffusion + generate + evaluate): bat\run_full_pipeline.bat

Generate samples from a trained model: bat\run_diffusion_multi.bat <model_path>

Evaluate caption adherence of a trained model: bat\evaluate_caption_adherence_multi.bat <model_path>

Running Python Scripts

Run the dataset/level JSON viewer python mm2_viewer_json.py <level.json>

Run the ASCII dataset browser python ascii_data_browser.py <dataset.json> <tileset.json>

Converting a .json to BCD python json_to_bcd.py <exact .json file location> --toost-compat --toost-compat is optional however it will make some custom levels appear in toost