How aloshdenny/reverse-synthid works
This repository is a research toolkit for analyzing SynthID image watermarks and generating watermark-reduced image variants, with a local drag-and-drop cleaner. The main workflow builds a spectral codebook from reference images, applies a selected bypass to input images, and can use manual Gemini detection tallies to recalibrate V4 carrier weights. Separate capabilities include watermark detection/extraction, investigation and analysis, and the desktop V3 workflow. Relationships involving unsampled modules are included only where the README or SOURCE INDEX supports them.
Main entry points: dissolve_batch.py Batch dissolver · gui.py Desktop cleaner · synthid_bypass_v4.py V4 consensus bypass · benchmark_extraction.py Extraction benchmark
Components 21
Watermark research
- Watermark investigation
watermark_investigation - Deep analysis
src/analysis/deep_synthid_analysis.py - Codebook discovery
src/analysis/synthid_codebook_finder.py
Detection and extraction
- Codebook extraction
src/extraction/synthid_codebook_extractor.py - Robust detection
src/extraction/robust_extractor.py - Extraction benchmark
src/extraction/benchmark_extraction.py
Bypass and cleaning
- Signature remover
src/extraction/watermark_remover.py - Spectral bypass V3
src/extraction/synthid_bypass.py - VAE regeneration
src/extraction/vae_regen.py - V4 consensus bypass
src/extraction/synthid_bypass_v4.py - Desktop cleaner
gui/gui.py
V4 calibration
- Codebook builder
scripts/build_codebook_v4.py - V4 spectral codebook
artifacts/spectral_codebook_v4.npz - Batch dissolver
scripts/dissolve_batch.py - Variant manifest
- Feedback calibration
scripts/calibrate_from_feedback.py
Other components
- Image user
- Reference images
- Gemini detection app
- Manual detection tally
- Cleaned image variants
Connections 25
- Image user → Desktop cleaner: selects images
gui/gui.py - Desktop cleaner → Spectral bypass V3: invokes bypass
gui/gui.py - Desktop cleaner → V4 consensus bypass: invokes bypass
gui/gui.py - Desktop cleaner → Cleaned image variants: saves images
gui/gui.py - Reference images → Codebook builder: provides references
README.md - Codebook builder → V4 consensus bypass: builds profiles
scripts/build_codebook_v4.py - Codebook builder → V4 spectral codebook: writes codebook
README.md - Image user → Batch dissolver: provides inputs
README.md - V4 spectral codebook → Batch dissolver: loads codebook
README.md - Batch dissolver → V4 consensus bypass: runs variants
scripts/dissolve_batch.py - Batch dissolver → Cleaned image variants: writes variants
README.md - Batch dissolver → Variant manifest: writes manifest
README.md - Variant manifest → Feedback calibration: joins feedback
scripts/calibrate_from_feedback.py - Manual detection tally → Feedback calibration: supplies tallies
scripts/calibrate_from_feedback.py - Feedback calibration → V4 spectral codebook: updates weights
README.md - Gemini detection app → Manual detection tally: provides detection results
README.md - V4 consensus bypass → VAE regeneration: runs regeneration
src/extraction/synthid_bypass_v4.py - Spectral bypass V3 → Robust detection: uses extractor
src/extraction/synthid_bypass.py - Signature remover → Robust detection: uses extractor
src/extraction/watermark_remover.py - Extraction benchmark → Robust detection: benchmarks detection
src/extraction/benchmark_extraction.py - Extraction benchmark → Signature remover: benchmarks removal
src/extraction/benchmark_extraction.py - Extraction benchmark → Codebook extraction: benchmarks extraction
src/extraction/benchmark_extraction.py - Image user → Watermark investigation: provides evidence
- Image user → Deep analysis: provides images
- Image user → Codebook discovery: provides images