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..
CORAL: Disentangling Latent Representations in Long-Tailed Diffusion
Authors: Esther Rodriguez, Monica Welfert, Samuel McDowell, Nathan Stromberg, Julian Camarena, Lalitha Sankar
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate that CORAL significantly improves both the diversity and visual quality of samples generated for tail classes relative to state-of-the-art methods. Through extensive experiments on several long-tailed datasets (CIFAR10-LT, CIFAR100-LT [15], Celeb A5 [16], Image Net-LT[17]), we demonstrate that CORAL significantly improves both the diversity and visual fidelity of tail-class samples, outperforming prior approaches. |
| Researcher Affiliation | Academia | Esther Rodriguez Arizona State University EMAIL Monica Welfert Arizona State University EMAIL Samuel Mc Dowell Arizona State University EMAIL Nathan Stromberg Arizona State University EMAIL Julian Antolin Camarena Arizona State University EMAIL Lalitha Sankar Arizona State University EMAIL |
| Pseudocode | Yes | Algorithm 1 CORAL Training Procedure |
| Open Source Code | Yes | The implementation code is available at https://github.com/Sankar Lab/coral-lt-diffusion. |
| Open Datasets | Yes | We evaluate CORAL on long-tailed (LT) datasets: CIFAR10-LT, CIFAR100-LT [28], Celeb A-5 [16], and Image Net-LT [17]. |
| Dataset Splits | Yes | For CIFAR10-LT and CIFAR100-LT, we simulate long-tailed distributions by applying an exponential decay to the class frequencies, controlled by an imbalance factor ρ {0.01, 0.001}. This results in the most frequent (head) class appearing 1/ρ times more often than the rarest (tail) class, with intermediate classes following an exponentially decreasing trend. Specifically for ρ = 0.01, CIFAR10-LT contains 12,406 images, with the first head class comprising 5,000 samples and the last tail class only 50. CIFAR100-LT has 10,847 images, with the head class containing 500 samples and the tail just 5. For evaluation, we generated 50k samples uniformly across all classes and compared against the balanced validation set containing 20k images as the real samples. |
| Hardware Specification | Yes | Training was run on NVIDIA A100 (80 GB SXM) and H100 GPUs. Experiments on CIFAR10-LT and CIFAR100-LT were conducted using NVIDIA A100 80GB SXM GPUs. Experiments on Celeb A-5 were run on NVIDIA H100 GPUs. Experiments on Image Net-LT were conducted using NVIDIA H100 GPUs. |
| Software Dependencies | No | Our implementation builds on the codebase from [5], with modifications to support contrastive latent regularization. We use a U-Net backbone with multi-resolution attention and dropout, consistent across all experiments. We use the Sup Con Loss implementation from [29]. |
| Experiment Setup | Yes | Experiments on CIFAR10-LT and CIFAR100-LT were conducted using NVIDIA A100 80GB SXM GPUs. Training took approximately 7 hours, and sampling required 8 hours. For DDPM, CBDM, and CORAL, the hyperparameters used were: a learning rate of 2 10 4, batch size of 128, Adam optimizer with default momentum parameters, dropout rate of 0.1, 150k training steps, and T = 1000 diffusion steps. For T2H, all settings remained the same except for the number of training steps, which was increased to 200k. Hyperparameters Table 3 summarizes the regularization hyperparameters and sampling guidance scale ω used for each method and dataset. The sub-tables correspond to DDPM (top left), CBDM (top right), T2H (bottom left), and CORAL (bottom right), respectively. For CORAL, the base contrastive weight, w, in (7) was set to 0.01. |