Research map / Image generation / Autoregressive and latent generation
LaRE^2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection
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 LaRE^2, a detector for diffusion-generated images that uses a single-step latent reconstruction error (LaRE) as a feature and refines CNN image features with it via error-guided spatial and channel modules.
Problem
Detecting diffusion-generated images generalizably and efficiently; existing reconstruction-based detectors (DIRE, SeDID) require multi-step DDIM inversion and reconstruction, which is slow and error-accumulating.
Approach
Compute the reconstruction error in latent space with a single denoising step at a fixed timestep (Monte Carlo averaged over noise samples) instead of full DDIM inversion-reconstruction, and use this error map to guide spatial attention and channel gating over CNN image features (EGRE).
Key results
- Average ACC 79.1% on GenImage across 8 generators, versus 67.2% for DIRE (reproduced), a gain of 11.9% ACC and 12.1% AP over the best SoTA.
- LaRE feature extraction is about 8x faster than DIRE.
- Ablation shows ESR adds 8.6%/5.3% ACC/AP and ECR adds 5.7%/2.4% over baseline; combined gain 11.4%/7.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ESR using LaRE instead of the image-only baseline (ablation) | LaRE2 · GenImage, average across 8 generators | accuracy: 73.1 -> 81.7 | Helped |
| add ESR using LaRE instead of the image-only baseline (ablation) | LaRE2 · GenImage, average across 8 generators | average precision: 92.3 -> 97.6 | Helped |
| add ECR using LaRE instead of the image-only baseline (ablation) | LaRE2 · GenImage, average across 8 generators | accuracy: 73.1 -> 78.8 | Helped |
| add ECR using LaRE instead of the image-only baseline (ablation) | LaRE2 · GenImage, average across 8 generators | average precision: 92.3 -> 94.7 | Helped |
| add both ESR and ECR using LaRE instead of the image-only baseline (ablation) | LaRE2 · GenImage, average across 8 generators | accuracy: 73.1 -> 84.5 | Helped |
| add both ESR and ECR using LaRE instead of the image-only baseline (ablation) | LaRE2 · GenImage, average across 8 generators | average precision: 92.3 -> 99.5 | Helped |
| use class-specific LaRE prompt instead of the generic prompt (ablation) | LaRE2 with ESR and ECR · GenImage, average across 8 generators | accuracy: +1.4% | Helped |
| use class-specific LaRE prompt instead of the generic prompt (ablation) | LaRE2 with ESR and ECR · GenImage, average across 8 generators | average precision: +0.3% | Helped |
| increase the noise ensemble size e (ablation) | LaRE2 · GenImage | accuracy | Helped |
| increase the noise ensemble size e (ablation) | LaRE2 · GenImage feature extraction | runtime | Hurt |
| choose sample step t outside the stable range [150, 300] (ablation) | LaRE2 · GenImage | performance | Hurt |
| use LaRE as the only input instead of image features (ablation) | generated-image detector · GenImage, average across 8 generators | accuracy: 73.1 -> 66.2 | Hurt |
Surprising findings
- A ResNet-20 trained on LaRE alone performs much worse (66.2 ACC) than the image-only baseline, showing LaRE is not sufficient alone.
Limitations
- Evaluation limited to GenImage; generalization to other datasets not shown in the main text.
- Detector trained separately per generator subset rather than a single unified model.
- Stated uncertainty about whether class-specific prompts are needed; generic prompt is used by default.
Experimental setup
Datasets: GenImage
Benchmarks: GenImage
Metrics: ACC AP
Baselines: CNNSpot Spec F3Net GramNet DIRE
Builds on: DIRE Latent Diffusion Models CLIP DDIM
Closest papers by meaning
- Aligned Datasets Improve Detection of Latent Diffusion-Generated Images (ICLR 2025 · similarity 0.88)
- FIRE: Robust Detection of Diffusion-Generated Images via Frequency-Guided Reconstruction Error (CVPR 2025 · similarity 0.87)
- DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection of Diffusion Generated Images (ICML 2024 · similarity 0.87)
- LMD: Faster Image Reconstruction with Latent Masking Diffusion (AAAI 2024 · similarity 0.87)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.86)
- RepLDM: Reprogramming Pretrained Latent Diffusion Models for High-Quality, High-Efficiency, High-Resolution Image Generation (NeurIPS 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.