Research map / Image generation / Autoregressive and latent generation
FIRE: Robust Detection of Diffusion-Generated Images via Frequency-Guided 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 FIRE, which detects diffusion-generated images by comparing autoencoder reconstruction error before and after removing learned mid-frequency bands, trained end-to-end with a ResNet-50 classifier.
Problem
Detecting images generated by diffusion models, especially generalizing to unseen generators and robustness to perturbations, where existing reconstruction-error detectors generalize poorly.
Approach
Observes that diffusion models struggle to reconstruct mid-band frequencies of real images; a learned Frequency Mask Refinement module isolates these bands, creates a pseudo-generated image by removing them, and a classifier uses the change in autoencoder reconstruction error (original vs. pseudo-generated) with end-to-end training.
Key results
- On DiffusionForensics (LSUN-Bedroom and ImageNet), FIRE reports an average of 100% AUC and ACC across tests.
- Trained on LAION + DALL-E 3, FIRE outperforms the next-best method FakeInversion by 3.2% AUC and 4.6% ACC; trained on LAION + Kandinsky 3, by 7.6% AUC and 7.3% ACC.
- FIRE is more robust to most perturbations (JPEG, cropping, blur, noise) than baselines, except under severe blur where FakeInversion is better.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FIRE instead of FakeInversion | Image detector · Newer generation models, LAION + DALL·E 3 training | AUC: +3.2% AUC | Helped |
| Use FIRE instead of FakeInversion | Image detector · Newer generation models, LAION + DALL·E 3 training | ACC: +4.6% ACC | Helped |
| Use FIRE instead of FakeInversion | Image detector · Newer generation models, LAION + Kandinsky 3 training | AUC: +7.6% AUC | Helped |
| Use FIRE instead of FakeInversion | Image detector · Newer generation models, LAION + Kandinsky 3 training | ACC: +7.3% ACC | Helped |
| Use preset frequency mask #5 instead of mask #1 (ablation) | FIRE · DALL·E 3 | AUC: 58.5 -> 78.3 | Helped |
| Use preset frequency mask #5 instead of mask #2 (ablation) | FIRE · DALL·E 3 | AUC: 73.4 -> 78.3 | Helped |
| Use preset frequency mask #5 instead of mask #3 (ablation) | FIRE · DALL·E 3 | AUC: 62.7 -> 78.3 | Helped |
| Use preset frequency mask #5 instead of mask #4 (ablation) | FIRE · DALL·E 3 | AUC: 76.5 -> 78.3 | Helped |
| Add the mask refinement loss Lmask (ablation) | FIRE · DALL·E 3 | AUC: 68.3 -> 78.3 | Helped |
| Add the mid-frequency reconstruction alignment loss Lmid rec (ablation) | FIRE · DALL·E 3 | AUC: 54.3 -> 78.3 | Helped |
Surprising findings
- Removing mid-band frequencies from real images lowers their reconstruction error, and the change is more pronounced for real than generated images.
Limitations
- Under severe blur, performance drops below FakeInversion because blur destroys high-frequency information, shrinking the real-generated reconstruction gap.
- Authors plan a more robust frequency refinement scheme to improve robustness.
Experimental setup
Datasets: DiffusionForensics LAION Midjourney-prompted generated images
Benchmarks: DiffusionForensics self-collected LAION + DALL-E 3 / Kandinsky 3 benchmark
Metrics: AUC ACC
Baselines: CNNDet AEROBLADE DIRE FakeInversion
Builds on: DIRE AEROBLADE FakeInversion ResNet-50
Closest papers by meaning
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.89)
- Warm Diffusion: Recipe for Blur-Noise Mixture Diffusion Models (ICLR 2025 · similarity 0.88)
- Boosting Diffusion Models with Moving Average Sampling in Frequency Domain (CVPR 2024 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.88)
- LaRE^2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection (CVPR 2024 · similarity 0.87)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 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.