Bypass does not work at all

[vstempire-ux]

SYNTHID BYPASS RESULTS

Pipeline: v3/aggressive
Success: True
PSNR: 38.82 dB
SSIM: 0.9925
Stages: pass_0(aggressive), pass_1(moderate), pass_2(gentle), anti_alias
version: v3_multi_res
strength: aggressive
passes: 3
pass_schedule: ['aggressive', 'moderate', 'gentle']
avg_luminance: 0.31854563569710925
profile_resolution: 1024x1024
exact_match: True
codebook_refs: 100b+100w+0r

Saved to: output1.png
(venv) sam@Macbook-Pro-M1 reverse-SynthID % python src/extraction/robust_extractor.py detect /Users/sam/reverse-SynthID/output1.png \
--codebook artifacts/codebook/robust_codebook.pkl

==================================================
ROBUST SYNTHID DETECTION RESULTS

Watermarked: True
Confidence: 0.5467
Correlation: -0.0043
Phase Match: 0.5607
Structure: 1.4828
Carrier Str: 225612.12
Multi-Scale: 0.0019

Detector returns 'True' for watermarked no matter what image or aggression level used in the bypass. So i guess this just doesn't work for some people? Or have i done something wrong?

[aloshdenny]

What model did you use to generate the images? If its NB Pro, likely lesser chance

[vstempire-ux]

What model did you use to generate the images? If its NB Pro, likely lesser chance

Yes it was all NB Pro, didnt realise this wasn't made to remove that, what is it more likely to work on?

[aloshdenny]

The codebook was built upon findings from the vanilla NB models.
Imagen 3 I believe.
If Im right, Imagen 4 is the Pro model.

[mrbeandev]

@aloshdenny This is the same root cause as #10, fixed by PR #18 (merged). The bypass itself was actually working — the detector just couldn't verify it because load_codebook() failed silently when given a .npz file instead of .pkl. After the fix, the bypass now shows before/after verification confirming watermark removal. Can this be closed?