Bypass does not work on images generated by Google Labs Flow

[jesufh]

I just discovered this projec, it's amazing. I tested it with images generated from Google Labs Flow (Nano Banana Pro model) with a 9:16 aspect ratio at 2K resolution. I followed the steps in the README, but the SynthID watermark wasn't removed from any of the images.

I'm not sure if I did something wrong or if there's an additional setting I need to adjust.

logs:

> python src/extraction/synthid_bypass.py bypass image_1.png outpot.png --codebook artifacts/spectral_codebook_v3.npz --strength aggressive
Warning: Could not load extractor: A load persistent id instruction was encountered,
but no persistent_load function was specified.
Codebook loaded: [1024x1024, 1536x2816]
  1024x1024: 100b+100w+0r
  1536x2816: 0b+0w+88r

============================================================
SYNTHID BYPASS RESULTS
============================================================
  Pipeline: v3/aggressive
  Success: True
  PSNR: 45.52 dB
  SSIM: 0.9968
  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: 0.4637169495417474
  profile_resolution: 1536x2816
  exact_match: False
  codebook_refs: 0b+0w+88r
============================================================
Saved to: outpot.png
> python src/extraction/robust_extractor.py detect outpot.png --codebook artifacts/codebook/robust_codebook.pkl

==================================================
ROBUST SYNTHID DETECTION RESULTS
==================================================
  Watermarked: False
  Confidence:  0.6442
  Correlation: -0.0012
  Phase Match: 0.3831
  Structure:   1.5083
  Carrier Str: 339086.09
  Multi-Scale: 0.0035
==================================================

[aloshdenny]

The spectral codebook has been constructed from images generated by Nano Banana
We would love contributions from the Pro model to improve the watermark ablation!

[jesufh]

Ooh, sure how can we contribute?

[aloshdenny]

Could you share a dataset of Pro images? Preferably a set of 100 pure black and 100 pure white images?

My workaround was to provide Nano Banana an image of pure white/pure black and then ask it to recreate it as it was.

[EfrainTorres]

Does it work with the API? I tried generating black images but it doesn't appear to have synth ID? When done through Gemini (and adds the watermark) it does?

[aloshdenny]

I'm not sure about the API. I had generated all the dataset images from Gemini - those have a watermark.

[EfrainTorres]

Question, I just noticed some of the comparison images are from SyncID Bypass repo. Why is that? Does this actually work? Happy to help with as many generations as needed for the codebook but.. that's not a good look. So far I haven't gotten it to work at all.