How dakshankarthic/ffdm works
This repository is a research codebase for detecting manipulated images and localizing edited regions, alongside baseline benchmark runners and an explicitly untrained global/local prototype. The main workflow validates and loads image/mask data, trains detector variants with segmentation and consistency objectives, then performs inference and benchmark scoring. The graph includes documented and indexed relationships; some experiment scripts and model internals are unsampled, so those links are omitted rather than inferred.
Main entry points: eval_benchmark_comparison.py Benchmark Evaluation · train.py Core Training · super_frame_v2_detector.py SuperFRAME v2 · forensic_encoder.py Forensic Encoder
Components 27
Data and Tiling
- Research Loader
controlled_localization/src/controlled_localization/data/research_loader.py - Manifest Schema
controlled_localization/src/controlled_localization/data/schema.py - Corrected Dataset
controlled_localization/src/controlled_localization/data/corrected_dataset.py - Tile Operations
controlled_localization/src/controlled_localization/data/tiles.py
Detection Models
- Segmentation Loss
controlled_localization/src/controlled_localization/losses/segmentation.py - Consistency Loss
controlled_localization/src/controlled_localization/losses/consistency.py - Base Detector
controlled_localization/src/controlled_localization/models/detector.py - Forensic Encoder
controlled_localization/src/controlled_localization/models/forensic_encoder.py - Bayar SRM Stem
controlled_localization/src/controlled_localization/models/bayar_srm_stem.py - Frame Detector
controlled_localization/src/controlled_localization/models/frame_detector.py - SuperFRAME
controlled_localization/src/controlled_localization/models/super_frame_detector.py - SuperFRAME v2
controlled_localization/src/controlled_localization/models/super_frame_v2_detector.py - DINO Encoder
controlled_localization/src/controlled_localization/models/dino_encoder.py - Cross-Attention Decoder
controlled_localization/src/controlled_localization/models/cross_attention_decoder.py
Training and Evaluation
- Core Training
controlled_localization/src/controlled_localization/train.py - Matched Pilot
controlled_localization/src/controlled_localization/train_matched_pilot.py - Dual-T4 Training
controlled_localization/src/controlled_localization/train_kaggle_dual_t4.py - Inference
controlled_localization/src/controlled_localization/infer.py - Decision Policy
controlled_localization/src/controlled_localization/policy.py - Benchmark Evaluation
controlled_localization/src/controlled_localization/eval_benchmark_comparison.py - Metrics
controlled_localization/src/controlled_localization/metrics.py
Baseline and Prototype
- Baseline Test Evaluation
baseline_work/code/evaluate_test_sets.py - VAAS Adapter
baseline_work/code/baseline_adapter.py - Global-Local Prototype
echoseam_research/model.py - Prototype Training
echoseam_research/train.py - Prototype Evaluation
echoseam_research/evaluate_pilot.py
Other components
- VAAS Model Service
Connections 35
- Research Loader → Manifest Schema: validates records
controlled_localization/src/controlled_localization/data/research_loader.py - Core Training → Research Loader: loads batches
controlled_localization/src/controlled_localization/train.py - Core Training → Base Detector: trains
controlled_localization/src/controlled_localization/train.py - Core Training → Segmentation Loss: computes loss
controlled_localization/src/controlled_localization/train.py - Core Training → Consistency Loss: computes consistency
controlled_localization/src/controlled_localization/train.py - Core Training → Decision Policy: uses policy
controlled_localization/src/controlled_localization/train.py - Core Training → Benchmark Evaluation: runs validation
controlled_localization/src/controlled_localization/train.py - Matched Pilot → Corrected Dataset: loads records
controlled_localization/src/controlled_localization/train_matched_pilot.py - Matched Pilot → Tile Operations: uses tiling
controlled_localization/src/controlled_localization/train_matched_pilot.py - Matched Pilot → SuperFRAME v2: trains
controlled_localization/src/controlled_localization/train_matched_pilot.py - Matched Pilot → Segmentation Loss: computes loss
controlled_localization/src/controlled_localization/train_matched_pilot.py - Dual-T4 Training → SuperFRAME v2: trains
controlled_localization/src/controlled_localization/train_kaggle_dual_t4.py - Dual-T4 Training → Tile Operations: uses tiling
controlled_localization/src/controlled_localization/train_kaggle_dual_t4.py - Dual-T4 Training → Segmentation Loss: computes loss
controlled_localization/src/controlled_localization/train_kaggle_dual_t4.py - Base Detector → Forensic Encoder: uses encoder
controlled_localization/src/controlled_localization/models/detector.py - Forensic Encoder → Bayar SRM Stem: uses stem
controlled_localization/src/controlled_localization/models/forensic_encoder.py - Frame Detector → Forensic Encoder: uses encoder
controlled_localization/src/controlled_localization/models/frame_detector.py - SuperFRAME → DINO Encoder: uses encoder
controlled_localization/src/controlled_localization/models/super_frame_detector.py - SuperFRAME → Forensic Encoder: uses encoder
controlled_localization/src/controlled_localization/models/super_frame_detector.py - SuperFRAME v2 → DINO Encoder: uses encoder
controlled_localization/src/controlled_localization/models/super_frame_v2_detector.py - SuperFRAME v2 → Forensic Encoder: uses encoder
controlled_localization/src/controlled_localization/models/super_frame_v2_detector.py - SuperFRAME v2 → Cross-Attention Decoder: uses decoder
controlled_localization/src/controlled_localization/models/super_frame_v2_detector.py - Inference → Tile Operations: tiles images
controlled_localization/src/controlled_localization/infer.py - Inference → Base Detector: runs model
controlled_localization/src/controlled_localization/infer.py - Inference → Decision Policy: applies policy
controlled_localization/src/controlled_localization/infer.py - Benchmark Evaluation → Inference: evaluates predictions
controlled_localization/src/controlled_localization/eval_benchmark_comparison.py - Benchmark Evaluation → Metrics: computes metrics
controlled_localization/src/controlled_localization/eval_benchmark_comparison.py - Benchmark Evaluation → Decision Policy: uses policy
controlled_localization/src/controlled_localization/eval_benchmark_comparison.py - Benchmark Evaluation → Frame Detector: evaluates variant
controlled_localization/src/controlled_localization/eval_benchmark_comparison.py - Benchmark Evaluation → SuperFRAME: evaluates variant
controlled_localization/src/controlled_localization/eval_benchmark_comparison.py - Benchmark Evaluation → SuperFRAME v2: evaluates variant
controlled_localization/src/controlled_localization/eval_benchmark_comparison.py - Baseline Test Evaluation → VAAS Adapter: loads and predicts
baseline_work/code/evaluate_test_sets.py - VAAS Adapter → VAAS Model Service: calls pipeline
baseline_work/code/baseline_adapter.py - Prototype Training → Global-Local Prototype: trains prototype
echoseam_research/train.py - Prototype Evaluation → Global-Local Prototype: evaluates prototype
echoseam_research/evaluate_pilot.py