Don't Play Favorites: Minority Guidance for Diffusion Models

Authors: Soobin Um, Suhyeon Lee, Jong Chul Ye

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on benchmark real datasets demonstrate that our minority guidance can greatly improve the capability of generating high-quality minority samples over existing generative samplers. We showcase that the performance benefit of our framework persists even in demanding real-world scenarios such as medical imaging, further underscoring the practical significance of our work.
Researcher Affiliation Academia Soobin Um, Suhyeon Lee & Jong Chul Ye KAIST, Daejeon, Republic of Korea {sum,suhyeon.lee,jong.ye}@kaist.ac.kr
Pseudocode No The paper describes its methods through prose and mathematical equations but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/soobin-um/minority-guidance.
Open Datasets Yes Datasets. Our experiments are conducted on six real benchmarks: four unconditional and two class-conditional datasets. For the unconditional settings, we employ Celeb A 64^2 (Liu et al., 2015), CIFAR-10 (Krizhevsky et al., 2009), and LSUN-Bedrooms 256^2 (Yu et al., 2015)... We use Image Net 64^2 and 256^2 (Deng et al., 2009)...
Dataset Splits No The paper mentions using specific subsets of real data for evaluating generated samples (e.g., "10K and 5K real samples yielding the highest Avgk NN values for Celeb A and CIFAR-10"), but these are for evaluation metrics, not explicit validation dataset splits for model training or tuning. There is no explicit description of a validation split for the main models used.
Hardware Specification Yes All these results were obtained using a single A100 GPU.
Software Dependencies No Our implementation is based on Py Torch (Paszke et al., 2019). While PyTorch is mentioned, a specific version number is not provided, nor are explicit versions for other libraries or dependencies.
Experiment Setup Yes For the number of minority classes L, we take L = 100 for the three unconditional natural image datasets and L = 25 for Image Net and CIFAR-10-LT. We use L = 50 for the brain MRI experiments... We swept over {1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 5.0, 6.0, . . . , 10.0} for the classifier scale w. We employ 250 timesteps to sample from the baseline DDPM... See Table 2 for explicit details.