🧋LATTE: Latent Trajectory Embedding for Diffusion-Generated Image Detection

Ana Vasilcoiu¹* Ivona Najdenkoska^1,2*, Zeno Geradts², Marcel Worring¹

¹University of Amsterdam, ²Netherlands Forensic Institute (NFI)

🌟 Highlights

• 🔬 Detection of generated images — A novel approach for detecting generated images by modeling the evolution of latent representations across the generative denoising process.
• 🌀 Latent Trajectory Modeling — Extracts multiple diffusion latents from Stable Diffusion into a trajectory sequence.
• 🔗 Latent–Visual Fusion — Aligns the extracted latents with visual semantics using ConvNeXt/CLIP vision encoders.
• 🧠 Robust & Generalizable — Outperforms AIDE and LaRE on GenImage, Chameleon and Diffusion Forensics, demonstrating both strong cross-generator and cross-domain performance.

🧩 Method Details

We construct the LATTE sequence by performing a single-step reconstruction for a selection of timesteps throughout the whole trajectory.

It encompasses two stages: (1) Latent–Visual Fusion, where the LATTE is fused with visual semantics through stacks of L cross-attention layers, and (2) Latent-Visual Classifier for average aggregation and output prediction.

🗂️ Project Structure

├── images
├──── # Folder with image resources
├── scripts
├──── # Folder with example scripts
├── clip_prompt_utils.py         # CLIP prompt tuning utilities prompt-tuning
├── dataset.py                   # Iterable dataset loader from cached latents
├── model.py                     # Model code for the different architectural configurations proposed
├── extract_latte.py             # Latent trajectory extraction from real/fake images
├── train.py                     # Distributed training script for LATTE classifier
├── test.py                      # Evaluation script for pretrained models
├── robustness.py                # Perturbation experiments and AP/accuracy visualization
├── heatmaps.py                  # Latent trajectory consistency analysis plotting 
└── README.md                    # You're here!

⚙️ Setup and Installation

Requirements

• Python 3.8+
• PyTorch 2.7.0+cuda12.6

The environment containing the rest of the required packages can be installed via:

conda env create -f environment.yml

🚀 How it works

1. Latent Extraction

Use extract_latte.py to preprocess and extract latent sequences for real and fake images:

python extract_latte.py \
  --real_folders /path/to/real \
  --fake_folders /path/to/fake \
  --cache_dirs /output/path \
  --data_size 224 224 \

2. Model Training

Train the LATTE classifier on cached latent sequences:

torchrun --nproc_per_node=4 train.py \
  --latent_dir_train /output/path \
  --latent_dir_validation /validation/path \
  --model_type "LatentTrajectoryClassifier" \
  --clip_type "convnext_base_in22k" \
  --epochs 20 \
  --process_latents_separately

3. Evaluation and Robustness Testing

Evaluate trained models and test robustness against perturbations:

python test.py \
  --checkpoint checkpoints/best_model.pth \
  --latent_dirs_test /path/to/test_chunks_adm /path/to/test_chunks_glide ... \
  --method_names ADM GLIDE ... \
  --model_type "LatentTrajectoryClassifier" \

python robustness.py \
  --checkpoint checkpoints/best_model.pth \
  --latent_dir /path/to/test_chunks \
  --model_type "LatentTrajectoryClassifier"

📊 Benchmarks

GenImage

Complete pairwise evaluation of detection performance across all 8 generators in the GenImage dataset. Each subplot corresponds to one detector - DIRE (left; baseline), LaRE (center; baseline), and LATTE (right; proposed) - and shows the accuracy(%) when training on the subset listed on the vertical axis and testing on the subset listed along the horizontal axis. Row- and column-averages summarize each method's cross-model generalization capabilities.

Chameleon

Results on the Chameleon benchmark highlight both the robustness of our approach and its effectiveness in generalizing across diverse visual domains.

Training set	AIDE (%)	LATTE (%)
SDv1.4	62.6	63.8
GenImage	65.8	68.3

DiffusionForensics

Results of a cross-domain generalization experiment where both models have been trained on the SDv1.4 subset of GenImage and tested on all generator subsets across the 3 dataset subsets of DiffusionForensics.

Subset	LaRE (%)	AIDE (%)	LATTE (%)
Bedroom	69.5	74.6	85.7
Celeba	90.0	75.5	91.1
Imagenet	89.9	76.2	91.1

📜 Citation

If you find the LATTE paper and code useful for your research and applications, please cite using this BibTeX:

@article{vasilcoiu2025latte,
  title={LATTE: Latent Trajectory Embedding for Diffusion-Generated Image Detection},
  author={Vasilcoiu, Ana and Najdenkoska, Ivona and Geradts, Zeno and Worring, Marcel},
  journal={arXiv preprint arXiv:2507.03054},
  year={2025}
}