Research map / Image generation / Autoregressive and latent generation
Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train the diffusion model on more images | DDPM · Celeb-A | similarity between generated images and training samples | Hurt |
| Train on the full dataset rather than a small subset | DDPM · Celeb-A | training-data copying | Hurt |
| Use split-product features instead of standard inner-product features | DINO image retrieval · 10 replication-detector datasets | average rank | Helped |
| Add CutMix augmentation during MoCo fine-tuning (ablation) | MoCo, ViT-B/16 · IN-Cutmix | mAP: 65.98 -> 55.74 | Hurt |
| Add CutMix augmentation during MoCo fine-tuning (ablation) | MoCo, ViT-B/16 · IN-Dif-Outpaint | mAP: 20.61 -> 46.96 | Helped |
| Use Stable Diffusion with captions containing key phrases associated with a memorized image | Stable Diffusion · Prompted image generation | replication occurrence: ≈ 20% | Helped |
| Use a LAION caption as the generation prompt rather than expecting reproduction of its source image | Stable Diffusion · LAION caption-source images | similarity to the caption source image | No clear effect |
| Match images with very high SSCD similarity rather than random LAION images | Stable Diffusion · LAION Aesthetics training-image matches | average number of training-set duplications: 3.1 -> 34.1 | Helped |
| Use a larger and more diverse training set | Diffusion models · Celeb-A and Oxford Flowers | training-data replication | Hurt |
| Use the ImageNet class-conditional LDM rather than the smaller-data diffusion models | ImageNet LDM · ImageNet | training-data replication: 0.6 | No clear effect |
Closest papers by meaning
- Image Copy Detection for Diffusion Models (NeurIPS 2024 · similarity 0.91)
- An Efficient Framework for Crediting Data Contributors of Diffusion Models (ICLR 2025 · similarity 0.88)
- UnlearnCanvas: Stylized Image Dataset for Enhanced Machine Unlearning Evaluation in Diffusion Models (NeurIPS 2024 · similarity 0.88)
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.87)
- On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.87)
- CommonCanvas: Open Diffusion Models Trained on Creative-Commons Images (CVPR 2024 · similarity 0.87)
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.