Research map / Image generation / GANs
Generative Model Perception Rectification Algorithm for Trade-Off between Diversity and Quality
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DMPRA, a search-evaluation-adjustment loop that finds low-density (minor group) samples in a generative model's perception space via extended neighborhood sampling, weights them informatively, and iteratively retrains to improve both diversity and quality.
Problem
Balancing diversity and quality in generative models when training data is imbalanced, causing abnormal perception of low-density (minor group) samples.
Approach
Identify low-density samples in a perception space built from a ResNet152 feature extractor trained on real and generated data, using an extended-neighborhood Chebyshev-style outlier threshold (3 std), then assign coefficient-of-variation-based informative sampling weights and dynamically re-train over multiple iterations.
Key results
- On StyleGAN2-ada, FID drops from 1.08 (iter 1) to 1.29 range-wise tables are inconsistent; reported improvements across iterations for LogcoshVAE FID from 131.72 to 97.66 over five iterations.
- In GAN comparison on CelebA-HQ, DMPRA-DRS achieves the best FID of 65.83 versus 66.21 for AdvLatGAN+DRS.
- Chest X-ray classification accuracy rises from 88.00% (StyleGAN2-ada augmentation) to 89.33% with DMPRA-trained augmentation; YOLOv5 detection accuracy 97.3% to 98.0% on a 50-image test set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DMPRA across five dynamic adjustments | LogcoshVAE on CelebA-HQ · CelebA-HQ | FID: 131.72±0.84 -> 97.66±0.06 | Helped |
| Apply DMPRA across five dynamic adjustments | StyleGAN2-ada on CelebA-HQ · CelebA-HQ | FID: 38.42±0.32 -> 23.08±0.09 | Helped |
| Apply DMPRA across five dynamic adjustments | Generative Flow on CelebA-HQ · CelebA-HQ | FID: 90.22±0.83 -> 87.13±0.19 | Helped |
| Apply DMPRA across five dynamic adjustments | Denoising Diffusion Model on CelebA-HQ · CelebA-HQ | FID: 98.25±0.68 -> 85.71±0.22 | Helped |
| Apply DMPRA with DRS instead of AdvLatGAN+DRS | StyleGAN2-ada on CelebA-HQ (5000 images) · CelebA-HQ (5000 images) | FID: 66.21 -> 65.83 | Helped |
| Apply DMPRA with DRS instead of AdvLatGAN+DRS | StyleGAN2-ada on CelebA-HQ (5000 images) · CelebA-HQ (5000 images) | PRDC:D: 0.2816 -> 0.3169 | Helped |
| Apply DMPRA with DRS instead of AdvLatGAN+DRS | StyleGAN2-ada on CelebA-HQ (5000 images) · CelebA-HQ (5000 images) | PRDC:C: 0.3170 -> 0.3356 | Helped |
| Apply DMPRA with DRS instead of AdvLatGAN+DRS | StyleGAN2-ada on CelebA-HQ (5000 images) · CelebA-HQ (5000 images) | BRISQUE: 19.58 -> 19.81 | Hurt |
| Use DMPRA-generated augmentation instead of StyleGAN2-ada augmentation | ResNet152 classification on Chest X-Ray Images · Chest X-Ray Images | accuracy: 88.00% -> 89.33% | Helped |
| Add 1000 DMPRA-generated images rather than 1000 images generated without DMPRA | YOLO v5 · Object detection test set | accuracy: 97.3% -> 98.0% | Helped |
| Remove informative weights from DMPRA (ablation) | LogcoshVAE · CelebA-HQ | FID | Hurt |
| Remove informative weights from DMPRA (ablation) | LogcoshVAE · CelebA-HQ | BRISQUE | Hurt |
Surprising findings
- Benefits of DMPRA are smaller for diffusion models and generative flows, which the authors attribute to their stronger inherent data perception.
Limitations
- Gains diminish for diffusion models and flows after several iterations
- Experiments use small subsets (e.g., 1000 or 5000 images, 50-image test set) and single-run-style reporting with limited statistical rigor
- Perception feature extractor and threshold choices are heuristic
Experimental setup
Datasets: CelebA-HQ AFHQ Chest X-Ray Images (Kermany et al. 2018)
Benchmarks: CelebA-HQ AFHQ Chest X-Ray Images
Metrics: FID PRDC BRISQUE classification accuracy detection accuracy
Baselines: LogcoshVAE StyleGAN2-ada Generative Flow Denoising diffusion model DRS Topk Dia-GAN AdvLatGAN LatentAugment
Builds on: Papadimitriou et al. 2003 (LOCI) Azadi et al. 2018 (Discriminator Rejection Sampling) Mo et al. 2019 DeVries et al. 2020
Closest papers by meaning
- Aligning Generative Denoising with Discriminative Objectives Unleashes Diffusion for Visual Perception (ICLR 2025 · similarity 0.86)
- Image Generation Diversity Issues and How to Tame Them (CVPR 2025 · similarity 0.85)
- Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model (CVPR 2025 · similarity 0.85)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.85)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.84)
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.