Skip to content

v4

Pre-release
Pre-release

Choose a tag to compare

View all tags
@aloshdenny aloshdenny released this 23 Apr 20:48
· 8 commits to main since this release
v4
302f7c7
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

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 × 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

File Purpose
src/extraction/synthid_bypass_v4.py Core codebook + all bypass methods
src/extraction/vae_regen.py SD-VAE re-generation stage (MPS / CUDA / CPU)
scripts/build_codebook_v4.py Build a codebook from the hierarchical dataset
scripts/dissolve_batch.py Run final / nuke presets over a folder
scripts/calibrate_from_feedback.py Update carrier weights from detection feedback

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
Assets 2
Loading