Research map / Image generation / Text-to-image diffusion
Understanding and Mitigating Copying in Diffusion Models
Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, Tom Goldstein
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 |
|---|---|---|---|
| Multiple captions (20 BLIP captions per image) at training | SD 2.1 finetuned · LAION-10k no duplication | dataset similarity (lower better): 0.638 -> 0.420 | Improved |
| Multiple captions at training | SD 2.1 finetuned · LAION-10k image + caption duplication | dataset similarity (lower better): 0.663 -> 0.470* | Improved |
| Random token replacement at inference | Stable Diffusion · LAION-10k image + caption duplication | dataset similarity (lower better): 0.663 -> 0.555 | Improved |
| Random number addition at inference | Stable Diffusion · LAION-10k image duplication | dataset similarity (lower better): 0.576 -> 0.497 | Improved |
| Random token replacement during training | SD 2.1 finetuned · LAION-10k image duplication | dataset similarity (lower better): 0.576 -> 0.645 | Hurt |
| Random token replacement during training | SD 2.1 finetuned · LAION-10k no duplication | dataset similarity (lower better): 0.638 -> 0.638 | No clear effect |
| Gaussian noise on text embeddings during training | SD 2.1 finetuned · LAION-10k no duplication | FID (lower better): 18.72 -> 20.63 | Hurt |
| Multiple captions during training | SD 2.1 finetuned · LAION-10k no duplication | FID (lower better): 18.72 -> 16.63 | Improved |
| Random token addition at inference | Stable Diffusion · LAION-10k no duplication | CLIP score: 30.52 -> 29.54 | Hurt |
| Use diverse (BLIP) captions instead of a fixed caption (ablation) | SD 2.1 finetuned on Imagenette · Imagenette | FID (lower better): 40.6 -> 17.74 | Improved |
| Partial duplication (resampled captions per duplicate) instead of full duplication, rising ddf (ablation) | SD 2.1 finetuned · LAION-10k / Imagenette | dataset similarity | Improved |
| Train the text encoder during finetuning (ablation) | SD 2.1 finetuned · Imagenette / LAION-10k | dataset similarity | Hurt |
Closest papers by meaning
- Image Copy Detection for Diffusion Models (NeurIPS 2024 · similarity 0.91)
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.91)
- On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Detecting, Explaining, and Mitigating Memorization in Diffusion Models (ICLR 2024 · similarity 0.90)
- Membership Inference on Text-to-Image Diffusion Models via Conditional Likelihood Discrepancy (NeurIPS 2024 · similarity 0.90)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.90)
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.
