Research map / Image generation / GANs
Procedural Level Generation with Diffusion Models from a Single Example
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains a latent DDPM on a single game level, using block2vec token embeddings and a fully convolutional, receptive-field-limited denoiser with large random crops, to generate playable Minecraft and Super Mario Bros levels of arbitrary size.
Problem
Procedural level generation with deep models is limited by scarce human-designed training levels, and GAN-based single-level generators produce incoherent artifacts such as floating blocks and suffer rising training costs with larger token vocabularies.
Approach
Replace one-hot token encoding with dense block2vec (word2vec-style) semantic embeddings, train a latent diffusion model (DDPM, x0-prediction) on them, and restrict the denoiser's receptive field by removing all up/downsampling and attention layers so it learns patch-level distributions from one example.
Key results
- On Minecraft levels, the method achieves the lowest TPKL-Div on Village (5.60 vs 8.04 WorldGAN and 8.09 TOAD-GAN 3D) and Plains (4.92 vs 6.77 and 13.68).
- On SMB level 1-2, the method reaches 87.8% completion and 73.33 user preference versus 86.5% and 66.92 for TOAD-GAN.
- Predicting x0 instead of noise reduces TPKL-Div from 19.08 to 7.49 and raises Levenshtein from 50492 to 94529 in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed diffusion model instead of World-GAN | Minecraft level generation · Minecraft village, swamp, plains, beach, and mineshaft | TPKL-Div: 8.04 -> 5.60; 11.19 -> 9.43; 13.68 -> 4.92; 14.75 -> 14.28; 15.76 -> 12.79 | Helped |
| Use the proposed diffusion model instead of GAN-based baselines | Minecraft level generation · Minecraft desert and beach | Levenshtein distance: 2657.54 -> 5087.05; 14221.77 -> 15101.45 | Mixed |
| Use the proposed diffusion model instead of TOAD-GAN | SMB level generation · SMB 1-2 and 4-2 | %Completed: 86.5 -> 87.8; 68.8 -> 70.2 | Helped |
| Use the proposed diffusion model instead of TOAD-GAN | SMB level generation · SMB 1-2 and 4-2 | User Preference: 66.92 -> 73.33; 66.05 -> 74.43 | Helped |
| Remove random cropping during training (ablation) | Ours · Minecraft sample generation | Visual quality | Hurt |
| Change the crop size (ablation) | Ours · Minecraft sample generation | TPKL-Div and Levenshtein distance: 5.34, 5.60, 5.35, 5.62, 6.03; 7881.32, 10324.40, 12782.03, 13007.54, 13382.25 | Mixed |
| Predict noise epsilon instead of x0 (ablation) | Ours · Single-example level generation | TPKL-Div: 7.49 -> 19.08 | Hurt |
| Predict noise epsilon instead of x0 (ablation) | Ours · Single-example level generation | Levenshtein distance: 94528.69 -> 50492.19 | Hurt |
Surprising findings
- Predicting the noise epsilon instead of x0 makes the single-example diffusion model hard to converge, yielding much worse TPKL-Div (19.08 vs 7.49).
Limitations
- Requires more training and inference time and compute than GAN-based methods.
- No conditional generation aligned with designer requirements yet.
- Results on TPKL-Div are mixed; Desert and Beach show the method is not uniformly best (e.g., Beach 14.28 vs 14.73 is close, Mineshaft 12.79 vs 15.76 better).
Experimental setup
Datasets: DREHMAL:PRIMORDIAL VGLC Super Mario Bros levels
Benchmarks: DREHMAL:PRIMORDIAL Minecraft levels VGLC Super Mario Bros levels
Metrics: TPKL-Div Levenshtein distance %Completed by A* agent user preference score
Baselines: WorldGAN TOAD-GAN
Builds on: DDPM word2vec WorldGAN SinGAN SinDiffusion
Closest papers by meaning
- Illuminating Mario Scenes in the Latent Space of a Generative Adversarial Network (AAAI 2021 · similarity 0.88)
- Moonshine: Distilling Game Content Generators into Steerable Generative Models (AAAI 2025 · similarity 0.87)
- Structured Generative Adversarial Networks (NeurIPS 2017 · similarity 0.86)
- Scaling Up Parameter Generation: A Recurrent Diffusion Approach (NeurIPS 2025 · similarity 0.86)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.84)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.