What's New in V4

V4 is a complete rebuild of the watermark-dissolving pipeline, adding multi-model support, a richer codebook, and a new all-in-one attack that defeats the SynthID detector on both gemini-3.1-flash-image-preview and nano-banana-pro-preview after six rounds of iterative adversarial refinement.

The Breakthrough: Round-06 All-in-One Attack ✓

Previous rounds (spectral subtraction, blog-guided bin targeting, denoise-residual phase, VAE re-generation) all failed. Round 06 works by stacking every SynthID failure mode the Gemini app itself documents into a single 7-stage pipeline:

1. VAE round-trip (Stable Diffusion sd-vae-ft-mse) — projects the image off the natural-image manifold the SynthID decoder was never trained against (Gowal et al. 2026 §6.1)
2. Elastic deformation — smooth low-frequency random warp field that fragments the watermark's spatial phase consensus, simulating the "collage effect" Gemini's own help text identifies as a detection failure
3. Global geometric combo — rotation + zoom + shift in one affine warp
4. Resize-squeeze — AREA downsample → LANCZOS upsample, erases sub-pixel watermark info
5. Color-contrast nudge — brightness / saturation / hue micro-shift
6. Residual-phase FFT subtraction — blog-universal + codebook-harvested carrier bins
7. JPEG chain + luma noise + bilateral filter

Two presets: final (balanced) and nuke (maximum strength). Output is visually indistinguishable from the original.

New: SpectralCodebookV4

• Multi-model: separate profiles for gemini-3.1-flash-image-preview and nano-banana-pro-preview
• Multi-color consensus: 6 solid-color backgrounds (black, white, blue, green, red, gray) per model per resolution — only bins with phase-locked across all colors are treated as carriers
• 7 resolutions ×ばつ 2 models = 14 profiles, covering all common Gemini output sizes
• Format v5: LZMA-compressed, int8 phase, uint8 mag/cw, sparse-zeroed below the 0.55 coherence threshold — 221 MB → 24 MB (89% reduction, lossless for the bypass)

⬇️ Download spectral_codebook_v4.npz from the release assets and place it in artifacts/.

New Files

Quickstart

pip install torch diffusers safetensors accelerate opencv-python numpy scipy
python scripts/dissolve_batch.py \
 --input ./watermarked/ \
 --output ./cleaned/ \
 --codebook artifacts/spectral_codebook_v4.npz \
 --model gemini-3.1-flash-image-preview \
 --strengths final nuke

What's Changed

• Fix multiple bugs in extraction pipeline by @hobostay in #13
• Add 1062 reference images at 2 new resolutions + expanded codebook by @BensonRen in #14
• Add 93 black reference images (gemini_black_nb_pro) by @mrbeandev in #17
• Fix codebook loading error when .npz path is passed to extractor by @mrbeandev in #18
• Migrate reference images to Hugging Face, remove from git by @mrbeandev in #22
• Fix detection: empirically verified carrier frequencies by @mrbeandev in #23

New Contributors

• @hobostay made their first contribution in #13
• @BensonRen made their first contribution in #14
• @mrbeandev made their first contribution in #17

Full Changelog: https://github.com/aloshdenny/reverse-SynthID/commits/v3

Contributors

• @BensonRen
• @mrbeandev
• @hobostay