Bot or Not — Denoising Trajectory Detector
Logistic-regression classifier on top of denoising-trajectory features extracted
with CLIP ViT-L/14 + Stable Diffusion v1.5. Reproduces the method from
Liang et al., "Denoising Trajectory Biases for Zero-Shot AI-Generated Image Detection" (NeurIPS 2025).
How it works
For each input image:
- Encode to SD v1.5 latent space (VAE).
- Add DDPM noise at timesteps
(50, 150, 300, 500, 800).
- Run one UNet denoising step per timestep with an empty-prompt embedding.
- Decode each denoised latent back to image space.
- Compute CLIP-cosine similarity between the original and each reconstruction.
This yields a 6-D feature vector [sim_mean, sim_t50, sim_t150, sim_t300, sim_t500, sim_t800],
which a logistic regression (class_weight='balanced', solver='lbfgs') classifies as AI / Real.
Training data
- AI images: 2,500 images generated by diffusion models (1024×1024).
- Real images: 2,500 images sampled from COCO 2017
train2017.
- 80/20 stratified split,
random_state=42.
Test metrics
Held-out test set: 1,000 images (500 Real, 500 AI), random_state=42.
| Metric | Value |
|---|
| Accuracy | 0.7940 |
| ROC AUC | 0.8679 |
| F1 | 0.7876 |
Per-class breakdown:
| Precision | Recall | F1 | Support |
|---|
| Real | 0.78 | 0.82 | 0.80 | 500 |
| AI | 0.81 | 0.76 | 0.79 | 500 |
Confusion matrix (rows = true, cols = predicted):
| Pred Real | Pred AI |
|---|
| True Real | 412 | 88 |
| True AI | 118 | 382 |
Usage
1from huggingface_hub import hf_hub_download
2import joblib, json
3# Or use the bundled inference module:
4# from inference import BotOrNotDetector
5# detector = BotOrNotDetector.from_pretrained("bezand/BoN1")
6# detector.predict("image.jpg")
A CUDA GPU is required for practical inference (~30s/image on a T4; CPU inference
is impractical because each prediction runs five SD denoising steps).
Files
classifier.joblib — trained sklearn.linear_model.LogisticRegression.scaler.joblib — StandardScaler fit on training features.config.json — feature-extractor config (timesteps, CLIP and SD model IDs).inference.py, feature_extractor.py — inference wrappers.
Limitations and biases
- Trained on a single AI-image source at fixed 1024×1024 resolution. Real images
(COCO) vary in size and content, which may bias the classifier toward
resolution/aspect-ratio cues rather than denoising-trajectory artefacts.
- Single-step denoising with an empty prompt; full multi-step trajectories may
give cleaner signal but were not used in training.
- Only tested against SD-family generators. Performance on other generators
(Midjourney, FLUX, autoregressive models) is unknown.
License
The trained classifier weights and StandardScaler are released under CC-BY-NC-4.0.
Inference also requires Stable Diffusion v1.5 (CreativeML Open RAIL-M) and
CLIP ViT-L/14, each governed by its own license.
Citation
1@inproceedings{liang2025denoising,
2 title = {Denoising Trajectory Biases for Zero-Shot AI-Generated Image Detection},
3 author = {Liang et al.},
4 booktitle = {NeurIPS},
5 year = {2025}
6}
