Procedural Level Generation with Diffusion Models from a Single Example

Authors: Shiqi Dai, Xuanyu Zhu, Naiqi Li, Tao Dai, Zhi Wang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our model is capable of generating stylistically congruent samples of arbitrary sizes compared to manually designed levels. It suits a wide range of level structures with fewer artifacts than GAN-based approaches. [...] Experiments Setup Datasets To evaluate the effectiveness and generality of our method, we choose two games that are widely used for procedural content generation: Minecraft and Super Mario Bros (SMB).
Researcher Affiliation Academia Shiqi Dai1, Xuanyu Zhu1, Naiqi Li1, Tao Dai3, Zhi Wang*1,2,4 1Shenzhen International Graduate School, Tsinghua University 2Tsinghua-Berkeley Shenzhen Institute, Tsinghua University 3College of Computer Science and Software Engineering, Shenzhen University 4Peng Cheng Laboratory
Pseudocode Yes Algorithm 1: Training on a single level x 1: xe EMB(x) 2: repeat 3: xe 0 Crop(xe) 4: t Uniform(1, . . . , T = 50) 5: ϵ N(0, I) 6: Take gradient descent step on: θ xe 0 exe 0,θ αtxe 0 + 1 αtϵ, t 2 7: until converged Conjointly, our training procedure is shown in Alg. 1.
Open Source Code Yes The source code is available at https://github.com/shiqi-dai/diffusioncraft.
Open Datasets Yes For Minecraft level generation, the dataset comes from a large handcrafted Minecraft world called DREHMAL:PRIMΩRDIAL (Awiszus, Schubert, and Rosenhahn 2021). The SMB level dataset comes from VGLC (Summerville et al. 2016), which is a corpus of video game levels in easily parseable formats.
Dataset Splits No The paper does not explicitly state specific training, validation, and test dataset splits with percentages or sample counts. It mentions training on 'each piece of level' and evaluating 'generated results' but not formal data partitioning for model development.
Hardware Specification Yes Experiments are performed on NVIDIA Tesla V100.
Software Dependencies No The whole framework is implemented by Py Torch. The paper does not provide specific version numbers for PyTorch or other software dependencies.
Experiment Setup Yes The latent diffusion model has a max time step T = 1000. We train it for 50000 iterations with an initial learning rate 5 10 4 and a batch size of 8.