How qixinyi1122-lucky/did works
DID detects AI-generated images by reconstructing inputs with guided diffusion, deriving first- and second-order residual features, training detectors on those features, and evaluating a two-branch fusion. The graph preserves the documented reconstruction outputs and the detector/fusion stages, plus the shared training and evaluation layers. Edges between sampled modules are supported by source excerpts or the SOURCE INDEX; residual and fusion relationships come from the README. Some runtime details, including detector inference, are unsampled, so uncertain links are omitted.
Main entry points: recon_metrics.py Feature extraction · train.py Detector training · evaluate.py Two-branch fusion · script_util.py Model factory
Components 28
Reconstruction
- Feature extraction
reconstruct/recon_metrics.py - Model factory
reconstruct/guided_diffusion/script_util.py - Diffusion sampler
reconstruct/guided_diffusion/gaussian_diffusion.py - Sampling schedules
reconstruct/guided_diffusion/respace.py - Diffusion U-Net
reconstruct/guided_diffusion/unet.py - Diffusion NN utilities
reconstruct/guided_diffusion/nn.py - Diffusion runtime
reconstruct/guided_diffusion/dist_util.py - Image loader
reconstruct/guided_diffusion/image_datasets.py - Diffusion logging
reconstruct/guided_diffusion/logger.py - Mixed precision
reconstruct/guided_diffusion/fp16_util.py - Timestep resampling
reconstruct/guided_diffusion/resample.py - Diffusion losses
reconstruct/guided_diffusion/losses.py - Residual feature roots
Detector
- Detector training
detector/train.py - Detector testing
detector/test.py
Shared ML
- Feature datasets
common/datasets.py - Experiment config
common/config.py - Model trainer
common/trainer.py - ResNet classifier
common/networks/resnet.py - Training utilities
common/utils.py - Early stopping
common/earlystop.py - LR warmup
common/warmup.py - Detector validation
common/eval.py
Fusion
- Two-branch fusion
fusion/evaluate.py - DIRE feature root
- DIFF feature root
- Detection result
Other components
- Dataset provider
Connections 25
- Dataset provider → Feature extraction: provides images
README.md - Feature extraction → Diffusion runtime: sets up device
reconstruct/recon_metrics.py - Feature extraction → Diffusion logging: configures logging
reconstruct/recon_metrics.py - Feature extraction → Image loader: loads images
reconstruct/recon_metrics.py - Feature extraction → Model factory: creates model
reconstruct/recon_metrics.py - Model factory → Diffusion sampler: builds diffusion
reconstruct/guided_diffusion/script_util.py - Model factory → Sampling schedules: builds schedule
reconstruct/guided_diffusion/script_util.py - Model factory → Diffusion U-Net: builds U-Net
reconstruct/guided_diffusion/script_util.py - Diffusion U-Net → Diffusion NN utilities: uses NN utilities
reconstruct/guided_diffusion/unet.py - Feature extraction → Residual feature roots: writes residual maps
README.md - Residual feature roots → Detector training: supplies features
README.md - Detector training → Experiment config: reads settings
detector/train.py - Detector training → Feature datasets: creates loader
detector/train.py - Detector training → Model trainer: trains model
detector/train.py - Model trainer → Training utilities: loads network
common/trainer.py - Training utilities → ResNet classifier: constructs classifier
common/utils.py - Detector training → Detector validation: validates model
detector/train.py - Detector validation → Feature datasets: creates loader
common/eval.py - Detector training → Early stopping: checks stopping
detector/train.py - Model trainer → LR warmup: uses scheduler
common/trainer.py - DIRE feature root → Two-branch fusion: provides first-order features
README.md - DIFF feature root → Two-branch fusion: provides second-order features
README.md - Two-branch fusion → Experiment config: reads settings
fusion/evaluate.py - Two-branch fusion → Detector validation: uses evaluation
fusion/evaluate.py - Two-branch fusion → Detection result: combines decisions
README.md