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..
Generative Data Augmentation via Diffusion Distillation, Adversarial Alignment, and Importance Reweighting
Authors: Ruyi An, haicheng huang, Huangjie Zheng, Mingyuan Zhou
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments We evaluate DAR-GDA on CIFAR-10 and Image Net-1K, comparing it with i) standard data-augmentation pipelines, ii) strong diffusion-based GDA baselines, and iii) a state-of-the-art GAN. The analysis proceeds in two stages. We benchmark it against conventional data augmentation methods and existing GDA baselines. We also assess its performance and adaptability with various underlying generative models. We additionally probe how DAR scales under dynamic augmentation on smaller datasets for its practical feasibility. |
| Researcher Affiliation | Collaboration | Ruyi An1 Haicheng Huang2 Huangjie Zheng3 Mingyuan Zhou1 1The University of Texas at Austin 2Shanghai Jiao Tong University 3Apple |
| Pseudocode | Yes | Algorithm 1 Generative data augmentation in supervised training with GDA-DAR framework. Input: Teacher diffusion model ̈̆̄; real dataset SP; hyperparameters: learning rates ̄̂̀,̄̂́,̄h, loss weights ̂1,̂2,̂aug, truncation threshold ̈; number of training iterations Ngen,Nclf. |
| Open Source Code | Yes | The source code is available at https://github.com/ruyianry/gda-dar. |
| Open Datasets | Yes | Datasets. We use the canonical train/val splits of CIFAR-10 [50] and Image Net-1K [17]; no external or test data are introduced at any stage, including teacher training, distillation, or classification. |
| Dataset Splits | Yes | Datasets. We use the canonical train/val splits of CIFAR-10 [50] and Image Net-1K [17]; no external or test data are introduced at any stage, including teacher training, distillation, or classification. [...] CIFAR-10 consists of 60,000 RGB images... By default, these are split into 5,000 training samples and 1,000 test samples per class... Image Net-1K is a dataset comprising 1.28 million images labeled across 1,000 object categories. |
| Hardware Specification | Yes | Experiments use NVIDIA A100-80GB GPUs-single-GPU for CIFAR-10, 8-GPU (DDP) for Image Net-1K implemented in Py Torch 2.1 [66]. For CIFAR-10, computations were performed on a single 80GB-A100 GPU. For other datasets, we utilized eight 80GB-A100 GPUs, employing the maximum possible batch size that is a power of two. |
| Software Dependencies | Yes | Experiments use NVIDIA A100-80GB GPUs-single-GPU for CIFAR-10, 8-GPU (DDP) for Image Net-1K implemented in Py Torch 2.1 [66]. On CIFAR-10, we set ̄ = 1.2 for Si DA and train with Adam (lr=1e-5) [46] optimizer. ... For Image Net-1K, ... using Adam W [58], and initial learning rate 3e-3. |
| Experiment Setup | Yes | Implementation Details. We instantiate the (D) and (A) steps of our DAR-GDA with two recent leading algorithms: CTM [43]... and Si DA [121]... On CIFAR-10, we set ̄ = 1.2 for Si DA and train with Adam (lr=1e-5) [46] optimizer. CTM is trained for 256 steps per batch with a student learning rate of 3e-4 and the discriminator learning rate 2e-3 (batch size 128). For Image Net-1K we distill EDM2XXL with Si DA across 8 A100-80GB GPUs: ̄ = 1.0, per-GPU batch size 64, gradient accumulation every 128 iterations using Adam (lr=5e-5) optimizer. For CIFAR-10 we train Res Net-18 [28] and VGG-16 [79]; for Image Net-1K we use Res Net-50 [28] and Vi T-S/16 [20] to evaluate the performance with and without the (R)eweighting component with self-normalization and ̈ = 1. CIFAR-10 models are trained for 300 epochs with batch size 128 using momentum SGD (lr=0.1). The hyperparameters ̂1 and ̂2 for the adversarial alignment objective follow the settings from prior work. On Image Net-1K, Res Net-50 is trained for 90 epochs with batch size 4096 and initial learning rate 1.6, while Vi T-S/16 is trained for 300 epochs with batch size 1024, Adam W [58], and initial learning rate 3e-3. We apply self-normalization and set ̈ to be 1 for obtaining r(x) for reweighting. |