Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Diffusion Models as Cartoonists: The Curious Case of High Density Regions

Authors: Rafał Karczewski, Markus Heinonen, Vikas Garg

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical findings reveal the existence of significantly higher likelihood samples that typical samplers do not produce, often manifesting as cartoon-like drawings or blurry images depending on the noise level. Curiously, these patterns emerge in datasets devoid of such examples. We also present a novel approach to track sample likelihoods in diffusion SDEs, which remarkably incurs no additional computational cost. Code is available at https://github.com/Aalto-Qu ML/high-density-diffusion
Researcher Affiliation Collaboration Rafał Karczewski, Markus Heinonen Aalto University EMAIL Vikas Garg Yai Yai Ltd and Aalto University EMAIL
Pseudocode Yes Algorithm 1 High density sampling 1: Input: Threshold t (0, T] 2: Initial x T N(0, σ2 T ID) 3: Sample xt pt(xt) eq. 6 or 7 4: y0 HD-ODE(t, 0, xt) eq. 18 5: Return y0
Open Source Code Yes Code is available at https://github.com/Aalto-Qu ML/high-density-diffusion
Open Datasets Yes As a demonstration, we train two versions of a diffusion model on CIFAR-10 (Krizhevsky et al., 2009), one with maximum likelihood training (Kingma et al., 2021; Song et al., 2021b) and one optimized for sample quality (Kingma & Gao, 2024). ... 2We used FFHQ-256 and Churches-256 models from github.com/yang-song/score_sde_ pytorch and Image Net-64 from github.com/NVlabs/edm
Dataset Splits No The paper mentions using CIFAR-10 for training and FFHQ-256 for testing, but does not specify the explicit splits (e.g., percentages, sample counts, or citations to predefined splits) used for these datasets during training, validation, or testing.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for conducting the experiments.
Software Dependencies No The paper mentions that the UNET parametrization uses an implementation from 'docs.kidger.site/equinox/examples/unet/', implying the use of the Equinox library. However, no specific version number for Equinox or any other key software dependencies (like Python, PyTorch, CUDA) is provided.
Experiment Setup Yes Specifically, these models are Variance Preserving (VP) SDEs with a linear log-SNR noise schedule and εparametrization... with hyperparameters: is biggan=True, dim mults=(1, 2, 2, 2), hidden size=128, heads=8, dim head=16, dropout rate=0.1, num res blocks=4, attn resolutions=[16]; trained for 2M steps, 128 batch size, and the adaptive noise schedule from Kingma & Gao (2024) with EMA weight 0.99.