Research map / Image generation / Autoregressive and latent generation
AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error
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 AEROBLADE, a training-free detector that flags latent diffusion model images by their low reconstruction error through the LDM's own autoencoder (LPIPS distance, minimum over AEs).
Problem
Detecting images generated by latent diffusion models (e.g., Stable Diffusion, Midjourney) amid growing visual disinformation; existing detectors need extensive training or costly diffusion processes and generalize poorly.
Approach
LDM-generated images lie on the manifold of the generator's autoencoder and are reconstructed with lower error than real images, so the minimum LPIPS distance between an image and its reconstructions from several known AEs gives a training-free detector that needs no denoising.
Key results
- AEROBLADE with LPIPS2 and minimum reconstruction error reaches AP of 0.979 to 0.999 across seven generators, nearly matching the trained Corvi et al. detector (AP 1.000 on all).
- Using the matching AE gives AP 0.991 to 0.999 with LPIPS2, and the minimum over AEs drops performance only slightly.
- DIRE performed very poorly (TPR@5%FPR 0.000 on all datasets), which the authors attribute to inconsistent JPEG vs PNG storage of its representations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AEROBLADE instead of the trained Corvi detector | generated-image detection · SD1.1 | AP: 1.000 -> 0.979 | No clear effect |
| use AEROBLADE instead of the trained Corvi detector | generated-image detection · SD1.5 | AP: 1.000 -> 0.978 | No clear effect |
| use AEROBLADE instead of the trained Corvi detector | generated-image detection · SD2.1 | AP: 1.000 -> 0.994 | No clear effect |
| use AEROBLADE instead of the trained Corvi detector | generated-image detection · KD2.1 | AP: 1.000 -> 0.999 | No clear effect |
| use AEROBLADE instead of the trained Corvi detector | generated-image detection · MJ4, MJ5, and MJ5.1 | AP: 1.000 -> 0.999; 1.000 -> 0.997; 1.000 -> 0.996 | No clear effect |
| take the minimum reconstruction error across multiple AEs instead of using a single AE (ablation) | AEROBLADE · SD1.1, SD1.5, SD2.1, KD2.1, MJ4, MJ5, and MJ5.1 | AP: 0.959; 0.957; 0.991; 0.996; 0.998; 0.993; 0.994 | Helped |
| use LPIPS2 instead of the standard LPIPS distance (ablation) | AEROBLADE with minimum reconstruction error · SD1.1, SD1.5, SD2.1, KD2.1, MJ4, MJ5, and MJ5.1 | AP: 0.979 vs 0.959; 0.978 vs 0.957; 0.994 vs 0.991; 0.999 vs 0.996; 0.999 vs 0.998; 0.997 vs 0.993; 0.996 vs 0.994 | Mixed |
| use deeper DDIM inversion and denoising reconstructions instead of AE-only reconstruction (ablation) | AEROBLADE · Stable Diffusion 1.5 and 2.1 | AP | Hurt |
| use LPIPS (VGG16) instead of alternative distance metrics (ablation) | AEROBLADE · SD1.1, SD1.5, SD2.1, KD2.1, MJ4, MJ5, and MJ5.1 | AP: 0.959; 0.957; 0.991; 0.996; 0.998; 0.993; 0.994 | Helped |
| use individual AE reconstruction errors instead of the minimum across AEs (ablation) | AEROBLADE · Midjourney images | AP: 0.999; 0.997; 0.994 | Mixed |
Surprising findings
- The DIRE detector, reported to generalize to Stable Diffusion 1.5, performed at chance-level or worse, which the authors trace to a data storage bias in its released representations.
- Midjourney images are detected almost perfectly using the Stable Diffusion 2 autoencoder, suggesting a shared AE architecture.
Limitations
- Best performance requires access to the autoencoder of the generator that produced the image.
- Low-complexity generated images such as logos are harder to detect.
- Robustness to perturbations such as JPEG compression depends on dataset and LPIPS layer; less robust than Corvi et al., which was trained on perturbed images.
Experimental setup
Datasets: LAION-5B Midjourney Discord images
Benchmarks: LAION-5B-derived generated sets (SD1.1, SD1.5, SD2.1, KD2.1) Midjourney Discord sets (MJ4, MJ5, MJ5.1)
Metrics: AP TPR@5%FPR
Baselines: Gragnaniello et al. Corvi et al. Ojha et al. DIRE SeDIDStat
Builds on: LPIPS latent diffusion models DIRE CLIP interrogator
Closest papers by meaning
- Aligned Datasets Improve Detection of Latent Diffusion-Generated Images (ICLR 2025 · similarity 0.88)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
- LaRE^2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection (CVPR 2024 · similarity 0.86)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.85)
- LMD: Faster Image Reconstruction with Latent Masking Diffusion (AAAI 2024 · similarity 0.85)
- SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.85)
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.