Research map / Image generation / Autoregressive and latent generation
AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction
Chao Wang, Zijin Yang, Yaofei Wang, Weiming Zhang, Kejiang Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add homogeneity calibration to loss ratio (ablation) | AEDR · SD2.1 target | attribution accuracy: 90.57% -> 99.66% | Improved |
| add homogeneity calibration (ablation) | AEDR · KD2.1 target | attribution accuracy: 78.35% -> 82.93% | Improved |
| add homogeneity calibration (ablation) | AEDR · FLUX target | attribution accuracy: 96.80% -> 96.65% | Hurt |
| add homogeneity calibration (ablation) | AEDR · VQDM target | attribution accuracy: 92.80% -> 90.85% | Hurt |
| use MSE instead of LPIPS as reconstruction loss (ablation) | AEDR · SD2.1 target | attribution accuracy: 90.70% -> 99.40% | Improved |
| use MSE instead of SSIM as reconstruction loss (ablation) | AEDR · SD2.1 target | attribution accuracy: 99.10% -> 99.40% | No clear effect |
| replace single gradient-based reconstruction loss with autoencoder double-reconstruction ratio | attribution method · Belonging vs other models | attribution accuracy: 70.4% -> 95.1% | Improved |
| replace gradient-based optimization with two autoencoder forward passes | attribution method · Runtime on FLUX | seconds per image: 30.64s -> 0.68s | Improved |
| apply AEDR to quantized autoencoders instead of continuous VAE | AEDR · VQ-VAE (VQDM) and MoVQ (KD2.1) | attribution accuracy: 96% or more on VAE -> 90.85% (VQ-VAE), 82.93% (MoVQ) | Hurt |
| AEDR vs LatentTracer on MoVQ model | AEDR · KD2.1 (MoVQ) | attribution accuracy: 93.13% -> 82.93% | Hurt |
Later papers that compared with this method
1 later accepted paper reports AEDR as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Data Provenance for Image Auto-Regressive Generation (ICLR 2026 · better on 8 of 8 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.86)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.85)
- HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025 · similarity 0.85)
- Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models (ICLR 2025 · similarity 0.85)
- LMD: Faster Image Reconstruction with Latent Masking Diffusion (AAAI 2024 · similarity 0.85)
- Epsilon-VAE: Denoising as Visual Decoding (ICML 2025 · similarity 0.84)
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.
