How debdev25/ledd works
LEDD is a research toolkit for detecting diffusion-generated images, with a dual-stream image classifier and evaluation/explanation workflows. The documented path is to prepare and leak-check an image archive, split and load it, train the staged detector, then evaluate and explain it; baseline comparisons and efficiency measurement are additional capabilities. Source excerpts show key model, data, training, and attribution dependencies. Some documented CLI capabilities are represented at their package directory because their implementation files were not sampled.
Main entry points: train.py Staged trainer · detector.py Dual-stream detector · evaluate.py Evaluation protocol · frequency.py Frequency stream
Components 22
Data pipeline
- Archive preparation
ledd/data/prepare.py - Leak check
ledd/data/leak_check.py - Generator splits
ledd/data/splits.py - Dataset and loaders
ledd/data/dataset.py - Image degradations
ledd/data/degradations.py
Detection model
- Dual-stream detector
ledd/models/detector.py - Spatial stream
ledd/models/spatial.py - Frequency stream
ledd/models/frequency.py - Spectrum and bands
ledd/models/fft.py - Stream fusion
ledd/models/fusion.py - Recorded attention
ledd/models/attention.py
Training
- Staged trainer
ledd/engine/train.py - Training objectives
ledd/losses
Evaluation and explanation
- Evaluation protocol
ledd/engine/evaluate.py - Classification metrics
ledd/engine/metrics.py - Attribution and faithfulness
ledd/explain - Attention relevance
ledd/explain/chefer.py - Efficiency measurement
ledd/efficiency.py - Threshold calibration
scripts/calibrate_threshold.py
Comparisons
- Baseline models
ledd/baselines/simple.py - External baselines
ledd/baselines/external.py
Other components
- Researcher
Connections 25
- Researcher → Archive preparation: builds archive
README.md - Researcher → Leak check: checks leakage
README.md - Researcher → Staged trainer: trains stages
README.md - Researcher → Evaluation protocol: runs protocol
README.md - Researcher → Attribution and faithfulness: requests explanations
README.md - Researcher → Efficiency measurement: measures cost
README.md - Researcher → Baseline models: compares models
README.md - Staged trainer → Generator splits: creates splits
ledd/engine/train.py - Staged trainer → Dataset and loaders: builds loaders
ledd/engine/train.py - Staged trainer → Dual-stream detector: builds model
ledd/engine/train.py - Staged trainer → Training objectives: uses objectives
ledd/engine/train.py - Dataset and loaders → Image degradations: applies augmentation
ledd/data/dataset.py - Dual-stream detector → Spatial stream: uses stream
ledd/models/detector.py - Dual-stream detector → Frequency stream: uses stream
ledd/models/detector.py - Dual-stream detector → Stream fusion: fuses streams
ledd/models/detector.py - Frequency stream → Spectrum and bands: extracts bands
ledd/models/frequency.py - Frequency stream → Recorded attention: encodes tokens
ledd/models/frequency.py - Stream fusion → Recorded attention: cross-attends
ledd/models/fusion.py - Evaluation protocol → Dataset and loaders: builds loaders
ledd/engine/evaluate.py - Evaluation protocol → Generator splits: loads partitions
ledd/engine/evaluate.py - Evaluation protocol → Classification metrics: computes scores
ledd/engine/evaluate.py - Attribution and faithfulness → Attention relevance: computes relevance
ledd/explain/__init__.py - Attention relevance → Recorded attention: reads attention
ledd/explain/chefer.py - Attribution and faithfulness → Classification metrics: evaluates faithfulness
ledd/explain/faithfulness.py - Baseline models → External baselines: offers adapters
ledd/baselines/__init__.py