Research map / Image generation / Autoregressive and latent generation
DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection of Diffusion Generated Images
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 DRCT, which trains AI-generated image detectors on diffusion-reconstructed real and fake images with a contrastive plus cross-entropy loss, and releases the DRCT-2M dataset for evaluating generalization to unseen diffusion models.
Problem
Generated image detectors trained on one diffusion model often fail to generalize to images from unseen diffusion models.
Approach
Create hard negatives by reconstructing real images through Stable Diffusion VAE-latent noising and DDIM denoising, then train detectors with a margin-based contrastive loss plus cross-entropy so they learn subtle diffusion artifacts.
Key results
- DRCT/UnivFD (SDv2 reconstruction) reaches 96.90% average accuracy on DRCT-2M-Wild and 91.35% on DRCT-2M.
- On GenImage, DRCT improves UnivFD average accuracy from 79.45% to 89.49% and Conv-B from 74.98% to 82.08% when trained on SDv1.4.
- Ablation on GenImage: contrastive loss adds 6.54% average accuracy, bringing Conv-B to 83.53%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add real reconstructed images to training (ablation) | Conv-B · GenImage | accuracy: increased by 6.55% | Helped |
| Add original SDv1.4 fake images to training with real and real reconstructed images (ablation) | Conv-B · GenImage | accuracy: increased by another 2.5% | Helped |
| Add fake reconstructed images to training (ablation) | Conv-B · GenImage | accuracy: additional 0.96% | Helped |
| Add contrastive loss to the BCE loss (ablation) | Conv-B with real, real reconstructed, fake, and fake reconstructed images · GenImage | accuracy: increase of 6.54% | Helped |
| Use SDv2 rather than SDv1 for reconstruction | DRCT · DRCT-2M | accuracy | Helped |
| Enhance UnivFD with DRCT | UnivFD · GenImage | accuracy: 79.45 -> 89.49 | Helped |
| Enhance UnivFD with DRCT | UnivFD · DRCT-2M-Wild | accuracy: 63.55 -> 96.90 | Helped |
Surprising findings
- Reconstructed real images alone as training negatives raise average GenImage accuracy by 6.55%, suggesting detectors partly overfit to semantic content rather than artifacts.
- GAN-generated images (e.g., BigGAN) see only modest gains from DRCT, which the authors attribute to differing generation artifacts.
Limitations
- Gains on GAN-generated images are smaller than on diffusion-generated images.
- Only evaluated on globally generated images; locally generated (inpainted) regions are not covered.
- Requires generating a large reconstruction dataset with diffusion models, adding cost.
Experimental setup
Datasets: MSCOCO DRCT-2M GenImage
Benchmarks: DRCT-2M DRCT-2M-Wild GenImage
Metrics: accuracy
Baselines: CNNSpot F3Net CLIP/RN50 GramNet De-fake DIRE Conv-B UnivFD
Builds on: DIRE UnivFD Stable Diffusion margin-based contrastive loss
Closest papers by meaning
- LaRE^2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection (CVPR 2024 · similarity 0.87)
- Classification Diffusion Models: Revitalizing Density Ratio Estimation (NeurIPS 2024 · similarity 0.87)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.86)
- Generalized Diffusion Detector: Mining Robust Features from Diffusion Models for Domain-Generalized Detection (CVPR 2025 · similarity 0.86)
- Aligned Datasets Improve Detection of Latent Diffusion-Generated Images (ICLR 2025 · similarity 0.86)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · 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.