Research map / Image generation / Autoregressive and latent generation
Image Copy Detection for Diffusion Models
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.
Builds D-Rep, a 40k-pair diffusion-replication image copy detection dataset with 6 human-labeled replication levels, and proposes PDF-Embedding, which supervises ViT representative vectors with KL divergence to smooth level PDFs.
Problem
Existing image copy detection models trained on hand-crafted transformations fail to identify content replicated by diffusion model generations, and no dataset or method exists for this graded replication setting.
Approach
Convert each annotated replication level into a smooth PDF over normalized levels, represent each image by a set of N+1 learned class-token vectors, and train with KL divergence between the PDF and the cosine-similarity distribution, so neighboring levels get smoothly related probabilities.
Key results
- Ours (Exp.) reaches 56.3% PCC and Ours (Gaussian) 24.0% RD on the D-Rep test set, beating the best protocol-driven baseline (Enlarging PCC: 54.4% PCC, 40.1% RD) and label smoothing (35.0% PCC).
- Best public baseline DINOv2 reaches 39.0% PCC and GPT-4V 47.3% PCC; current ICD models SSCD and BoT reach 29.1% and 35.6% PCC.
- Evaluating six diffusion models against LAION-Aesthetics V2 6+ gallery, replication ratios range from 10.91% (SDXL) to 20.21% (Midjourney).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PDF supervision instead of directly optimizing PCC | PDF-Embedding on D-Rep · D-Rep test | PCC (%): 54.4 -> 56.3 | Helped |
| Use PDF supervision instead of directly optimizing PCC | PDF-Embedding on D-Rep · D-Rep test | RD (%): 40.1 -> 24.0 | Helped |
| Use PDF supervision instead of directly optimizing RD | PDF-Embedding on D-Rep · D-Rep test | PCC (%): +16.0% | Helped |
| Use PDF supervision instead of directly optimizing RD | PDF-Embedding on D-Rep · D-Rep test | RD (%): -4.1% | Helped |
| Replace one-hot supervision with exponential PDF supervision (ablation) | PDF-Embedding on D-Rep · D-Rep test | PCC (%): +18.7% | Helped |
| Replace one-hot supervision with exponential PDF supervision (ablation) | PDF-Embedding on D-Rep · D-Rep test | RD (%): 43.3 -> 25.6 | Helped |
| Replace label smoothing with exponential PDF supervision (ablation) | PDF-Embedding on D-Rep · D-Rep test | PCC (%): +21.3% | Helped |
| Replace label smoothing with exponential PDF supervision (ablation) | PDF-Embedding on D-Rep · D-Rep test | RD (%): -10.5% | Helped |
| Use exponential rather than Gaussian PDF supervision (ablation) | PDF-Embedding on D-Rep · D-Rep test | PCC (%): 53.7 -> 56.3 | Helped |
| Use Gaussian rather than exponential PDF supervision (ablation) | PDF-Embedding on D-Rep · D-Rep test | RD (%): 25.6 -> 24.0 | Helped |
| Use exponential rather than Gaussian PDF supervision (ablation) | PDF-Embedding on other diffusion models · D-Rep test, generalization | average similarity: 0.754 -> 0.814 | Helped |
Surprising findings
- Existing ICD models trained for transformation invariance (SSCD, BoT) perform worse than general vision-language and supervised models on diffusion replication detection.
- GPT-4V Turbo, used zero-shot, costs roughly $7,800 to compare one image against a one-million-image database through the API.
Limitations
- Replication labels are manually annotated by ten labelers with voting, so labels are subjective and costly to obtain (one month of labeling).
- Failure cases include common elements (grass, buildings) or high-level semantic similarity being mistaken for replication.
- Generalization evaluation uses only 100 manually confirmed pairs per diffusion model, and the PDF is discrete despite the continuous formulation.
Experimental setup
Datasets: D-Rep LAION-Aesthetics V2 6+ DiffusionDB ImageNet
Benchmarks: D-Rep test set generalization across diffusion models (Midjourney, DALL·E 2, DeepFloyd IF, New Bing, SDXL, GLIDE)
Metrics: PCC RD similarity score
Baselines: CLIP BLIP SLIP GPT-4V DINOv2 SSCD BoT ASL CNNCL EfNet SimCLR MAE
Builds on: DeiT SSCD BoT Sentence Transformer
Closest papers by meaning
- CommonCanvas: Open Diffusion Models Trained on Creative-Commons Images (CVPR 2024 · similarity 0.88)
- Membership Inference on Text-to-Image Diffusion Models via Conditional Likelihood Discrepancy (NeurIPS 2024 · similarity 0.88)
- Diffusion Models as Cartoonists: The Curious Case of High Density Regions (ICLR 2025 · similarity 0.87)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.87)
- Towards Black-Box Membership Inference Attack for Diffusion Models (ICML 2025 · similarity 0.87)
- Diffusing Differentiable Representations (NeurIPS 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.