Research map / Image generation / Autoregressive and latent generation
Will Large-scale Generative Models Corrupt Future Datasets?
Ryuichiro Hataya, Han Bao, Hiromi Arai
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace 80% of ImageNet training images with Stable Diffusion images | ResNet-50 · ImageNet val | top-1 accuracy: 75.7 -> 65.3 | Hurt |
| replace 20-40% of ImageNet images with generated images | ResNet-50 · ImageNet val | top-1 accuracy: 75.7 -> 74.5 (20%), 72.6 (40%) | Hurt |
| use real WebVision images instead of generated images for contamination (ablation) | ResNet-50 · ImageNet val | top-1 accuracy: IN/SD-80% 65.3 vs IN/WV-80% 68.3 | Improved |
| replace COCO captions images with generated images | BLIP captioner · COCO test | CIDEr: 1.335 -> 1.279 (80% generated) | Hurt |
| replace ImageNet images with generated images in unconditional diffusion training | IDDPM · ImageNet-64 | Recall@5: 0.644 -> 0.364 (100% generated) | Hurt |
| replace ImageNet images with generated images in diffusion training | IDDPM · ImageNet-64 | FID: 14.9 -> 11.0 (40%), 16.9 (100%) | Mixed |
| use more complex, diverse prompts to generate SD-ImageNet (ablation) | ResNet-50 · ImageNet val | top-1 accuracy: 15.7 -> 18.6 | No clear effect |
| contaminate with generated images | ResNet-50 · ImageNet-R | accuracy: 36.7 -> 30.0 (80% generated) | Hurt |
| append generated images to full ImageNet instead of replacing | ResNet-50 · ImageNet val | top-1 accuracy: 75.7 -> 75.7 (+20%), 75.0 (+80%) | No clear effect |
| fill missing 95% of ImageNet with generated images when only 5% real is available | ResNet-50 · ImageNet val | top-1 accuracy: 44.7 -> 52.1 | Improved |
| fill missing COCO data with generated captions images | BLIP captioner · COCO test (5% real) | CIDEr: 1.290 -> 1.220 | Hurt |
| use self-supervised MAE pretraining instead of supervised training with generated data | ViT-B · ImageNet val (linear probe) | accuracy: SD-ImageNet 38.8 vs ImageNet 43.9 | Improved |
Closest papers by meaning
- Dealing with Synthetic Data Contamination in Online Continual Learning (NeurIPS 2024 · similarity 0.89)
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 2024 · similarity 0.88)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.87)
- Would Deep Generative Models Amplify Bias in Future Models? (CVPR 2024 · similarity 0.87)
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.86)
- Anomaly Score: Evaluating Generative Models and Individual Generated Images based on Complexity and Vulnerability (CVPR 2024 · similarity 0.86)
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.
