DiffusionPrint: Learning Generative Fingerprints for Diffusion-Based Inpainting Localization

This is the official repository for the paper: DiffusionPrint: Learning Generative Fingerprints for Diffusion-Based Inpainting Localization Paschalis Giakoumoglou, Symeon Papadopoulos CVPR Workshop on AI for Media Integrity and Security (AIMS), 2026

[Paper] [Dataset]

---

Overview

DiffusionPrint learns per-generator fingerprints from 64x64 image patches using a MoCo-style contrastive framework built on a DnCNN backbone. The learned features replace the noiseprint++ stream in TruFor for pixel-level localization of diffusion-based inpainting artifacts.

The training pipeline uses hard negative mining over a feature queue and supports an optional classification head for generator identification. At inference, the DiffusionPrint backbone is plugged into TruFor as a drop-in feature extractor; see the TruFor/ directory for evaluation code.

---

Requirements

conda create -n diffusionprint python=3.11
conda activate diffusionprint
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txt

---

Dataset

The DiffusionPrint patch dataset is hosted on HuggingFace: giakoupg/diffusionprint_dataset.

It consists of 64x64 RGB patches extracted from real and diffusion-inpainted images across three generators (SD2, SDXL, Flux). Each patch is paired with at least one positive patch (same image region, different diffusion run) for contrastive training.

Download via HuggingFace datasets:

from datasets import load_dataset
ds = load_dataset("giakoupg/diffusionprint_dataset", split="train")

Or use the provided dataset loader directly in training (see dataset.py):

from dataset import DiffusionPrintDatasetHF
dataset = DiffusionPrintDatasetHF(repo_id="giakoupg/diffusionprint_dataset")

For training with the local .dat memmap format (used internally at MeVer), use DiffusionPrintDataset from dataset.py with --train_csv and --patches_dir.

---

Training

python train_diffusionprint.py \
    --train_csv /path/to/patches.csv \
    --patches_dir /path/to/patches.dat \
    --mode projector \
    --out_channels 64 \
    --projector_hidden_dim 128 \
    --projection_dim 128 \
    --temperature 0.07 \
    --batch_size 512 \
    --lr 0.001 \
    --epochs 50 \
    --top_k 512 \
    --cls_lambda 0.5 \
    --num_classes 2 \
    --augmentations geometric \
    --save_dir ./ckpt

Key arguments:

Argument	Description
--mode	projector (GAP + MLP + cosine) or flatten_projector
--out_channels	DnCNN output channels
--top_k	Number of hard negatives mined per batch from queue
--cls_lambda	Weight of auxiliary classification loss (0 to disable)
--augmentations	none, geometric, mild, or strong
--exclude_generators	Exclude one or more generators, e.g. --exclude_generators flux
--input_highpass	Apply Gaussian high-pass to all inputs before backbone
--noiseprint_weights	Initialize DnCNN from noiseprint++ weights
--resume	Resume from a checkpoint

Checkpoints and training logs (params.json, losses.csv) are saved under --save_dir/<run_name>/.

---

Pretrained Weights

Pretrained DiffusionPrint checkpoints will be released shortly. Place downloaded weights under pretrained/diffusionprint/ and pass the path via --noiseprint_weights or --pretrained when training, or configure the path in TruFor/config.py for evaluation.

---

Evaluation

Evaluation is handled by the TruFor integration in the TruFor/ subdirectory. See TruFor/README.md for setup and usage instructions.

Quick start:

python TruFor/evaluate.py \
    --image example.png \
    --mask example_mask.png \
    --model_path TruFor/weights/trufor_diffusionprint_tgif/ckpt.pth \
    --exp TruFor/trufor_diffusionprint.yaml \
    --save_maps --maps_dir ./maps

---

Repository Structure

diffusionprint/
├── builders/
│   ├── DnCNN.py              # DnCNN backbone
│   ├── diffusionprint.py     # MoCo contrastive model with hard negative mining
│   └── __init__.py
├── dataset/
│   └── dataset.py            # Patch dataset loaders (memmap and HuggingFace)
├── TruFor/                   # TruFor integration (localization + evaluation)
│   └── README.md
├── train_diffusionprint.py   # Training entry point
└── requirements.txt

---

Citation

If you use this code or dataset, please cite:

@inproceedings{giakoumoglou2026diffusionprint,
  title={DiffusionPrint: Learning Generative Fingerprints for Diffusion-Based Inpainting Localization},
  author={Giakoumoglou, Paschalis and Papadopoulos, Symeon},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8654--8664},
  year={2026}
}

---

Contact

For questions or inquiries, please contact giakoupg@iti.gr