Research map / Image generation / Autoregressive and latent generation
Data Provenance for Image Auto-Regressive Generation
Bihe Zhao, Louis Kerner, Michel Meintz, Tameem Bakr, Franziska Boenisch, Adam Dziedzic
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 calibration to the encoder reconstruction loss (ablation) | EncLoss on VAR · ImageNet non-belonging | TPR@1%FPR (%): 17.0 -> 100.0 | Improved |
| add calibration to the encoder reconstruction loss (ablation) | EncLoss on RAR · ImageNet non-belonging | TPR@1%FPR (%): 22.6 -> 98.2 | Improved |
| add calibration to the encoder reconstruction loss (ablation) | EncLoss on Infinity · ImageNet non-belonging | TPR@1%FPR (%): 0.3 -> 0.0 | Hurt |
| finetune inverse decoder instead of original encoder (ablation) | combined loss on RAR · ImageNet non-belonging | TPR@1%FPR (%): 6.2 -> 100.0 | Improved |
| finetune inverse decoder instead of original encoder (ablation) | combined loss on VQDiff · LAION non-belonging | TPR@1%FPR (%): 33.2 -> 99.4 | Improved |
| add optimized token search to quantization loss (ablation) | Quantization loss on VAR · ImageNet non-belonging | TPR@1%FPR (%): 0.4 -> 95.0 | Improved |
| combine quantization loss with encoder loss (ablation) | Quantization loss on VAR · ImageNet non-belonging | TPR@1%FPR (%): 0.4 -> 100.0 | Improved |
| combine quantization and encoder losses (ablation) | Quantization loss on Infinity · ImageNet non-belonging | TPR@1%FPR (%): 99.4 -> 0.0 | Hurt |
| finetune inverse decoder with augmentations (ablation) | Quantization loss on RAR · Noise (0.05) attack | TPR@1%FPR (%): 60.4 -> 87.8 | Improved |
| finetune inverse decoder with augmentations (ablation) | Quantization loss on RAR · Contrast (2.0) attack | TPR@1%FPR (%): 45.7 -> 91.1 | Improved |
| increase optimization iterations from 100 to 1200 (ablation) | Optimized token search on VAR · ImageNet non-belonging | TPR@1%FPR (%): 87.5 -> 95.0 | Improved |
| increase optimization iterations from 1200 to 1600 (ablation) | Optimized token search on VAR · ImageNet non-belonging | TPR@1%FPR (%): 95.0 -> 92.2 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AEDR — AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction (AAAI 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
- Watermarking Autoregressive Image Generation (NeurIPS 2025 · similarity 0.87)
- Privacy Attacks on Image AutoRegressive Models (ICML 2025 · similarity 0.87)
- On the Robustness of Watermarking for Autoregressive Image Generation (NeurIPS 2026 · similarity 0.86)
- A Locally Tokenized Generative Model for Robust Time-Series Watermarking (NeurIPS 2026 · similarity 0.85)
- BitMark: Watermarking Bitwise Autoregressive Image Generative Models (NeurIPS 2025 · similarity 0.85)
- Training Data Provenance Verification: Did Your Model Use Synthetic Data from My Generative Model for Training? (CVPR 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.
