Desynth

A tool & pipeline for removing OpenAI & Google's SynthID watermark from images.
This project is intended solely for research, education, and authorized evaluation.

Results

Edge Mode: Gaussian Mode (Default):

Metrics

All scores are input image vs output via the included compare.py.

Against other research (Gemini/Google, 2752×1536)

Our own test image (GPT Image 2.0/OpenAI, 1460×1078)

• Edge mode trades a small amount of measurable detail for better perceptual shape continuity at contours.

How it works

flowchart TD
    A["Original.png"]
    B["Qwen-Image GGUF Q4<br>+ Lightning 4-step LoRA"]
    B_note["2 Lightning steps\n(strength 0.25)"]
    C["Frequency-domain restore<br>low_clean + high_orig"]
    C_note["Gaussian split, sigma=1.95<br>low band from clean<br>high band from Original"]
    D["[output]\nNAME_desynth_r1.95.png"]

    A --> B
    B -- "clean: no SynthID, blurry-ish" --> C
    C --> D
    B -.- B_note
    C -.- C_note

    classDef note fill:#f6f8fa,stroke:#d0d7de,color:#57606a;
    class B_note,C_note note;

Usage

Download the two model files into the repo root:

Run

python desynth.py                          # processes original.png
python desynth.py path\to\image.png        # processes any input

Output: out/<name>_desynth_s8_d0.250_p1_r1.95.png. Random seed per run by default. First run downloads ~250 MB of VAE + configs from Hugging Face and caches them.

Flags

Quality check

python compare.py Original.png out\<output>.png

Prints PSNR, SSIM (full + low/high band), MAE, MSE, and per-channel histogram correlation.

Files

Hardware Requirements

Tested on Windows 10, RTX 5060 Ti 8 GB, 32 GB DDR4 RAM. Sequential CPU offload is required with < 12GB VRAM

Known limitations

• Lightning's 4-step distillation is the source of most of the residual drift.
  (Dropping it for proper 20+ step sampling would likely tighten metrics further at the cost of 5x longer runs.)

Credits

Baseline workflow and watermark hypothesis from 00quebec/Synthid-Bypass.
This pipeline reimplements the core idea in plain Python without ComfyUI, ControlNet, or the face-detail path, and replaces the heavier redraw with a two-step minimum denoise + frequency-domain restore.