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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Principled Long-Tailed Generative Modeling via Diffusion Models

Authors: Pranoy Das, Kexin Fu, Abolfazl Hashemi, Vijay Gupta

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform empirical validation of our theoretical findings with the widely used CIFAR10 dataset in the domain of image synthesis, specifically its long-tailed versions CIFAR10LT. The construction of CIFAR10LT follows from [5], where the size decreases exponentially with its class label index according to the imbalance factor imb = 0.01. We also perform experiments on synthetic dataset such as Gaussian Mixture Model and include them in Appendix C.
Researcher Affiliation Academia Pranoy Das, Kexin Fu, Abolfazl Hashemi, Vijay Gupta School of Electrical and Computer Engineering, Purdue University, W. Lafayette, IN, 47906 EMAIL
Pseudocode Yes Algorithm 1 Individual Gradient Descent(IGD)
Open Source Code Yes The code is available at https://github.com/pranoydas51/IGD-ML
Open Datasets Yes We perform empirical validation of our theoretical findings with the widely used CIFAR10 dataset in the domain of image synthesis, specifically its long-tailed versions CIFAR10LT. The construction of CIFAR10LT follows from [5]
Dataset Splits No The construction of CIFAR10LT follows from [5], where the size decreases exponentially with its class label index according to the imbalance factor imb = 0.01. We also perform experiments on synthetic dataset such as Gaussian Mixture Model and include them in Appendix C.
Hardware Specification Yes The numerical experiments were conducted on a Mac Book Air (2023) and Gilbreth. Gilbreth has heterogeneous hardware comprising of Nvidia V100, A100, A10, and A30 GPUs in separate sub-clusters. All the nodes are connected by 100 Gbps Infiniband interconnects. We used sub-cluster B with 16 nodes, 24 cores per node, 192 GB memory per node, 3 A30 (24 GB) per node.
Software Dependencies No The Neural network Architecture employed is U-net as in [23]. To be able to make direct comparisons to DDPM and a rudimentary comparison to CBDM, we modify the code of CBDM and employ error networks for mutual learning (individual gradient descent) instead of score networks as above.
Experiment Setup Yes We run both CBDM and Individual Gradient Descent(IGD) for Ttrain = 60k training steps. We generate 15k samples per class and make the comparison at the 60k training step mark. Implementation Details We employ the random feature model with the width of network m = 16, learning rate ητ = 10 4, τ, Ttrain = 5000 is fixed for Adam optimizer. We set λ(t) = σt, ω(t) = et, total number of training samples is 50. Figure 4: FID and IS Scores for Different β and η Values