DeSynth

KO | EN

gitlite — search

#python #php #dotfiles #arduino #linux #android #javascript #nodejs #docker #atom #ios #bot

★ 24 Open GitHub ↗

DeSynth removes the SynthID invisible watermark from both OpenAI/Google generated images. (SynthID Bypass)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 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:
![reference vs desynth (edge mode)](https://raw.githubusercontent.com/0xROOTPLS/DeSynth/main/out/comparison_Original_desynth_s8_d0.250_p1_r1.95_edge.png)
Gaussian Mode (Default):
![reference vs desynth (gaussian mode)](https://raw.githubusercontent.com/0xROOTPLS/DeSynth/main/out/comparison_Original_desynth_s8_d0.250_p1_r1.95.png)

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

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

| metric                 | our method | [competitor](https://github.com/00quebec/Synthid-Bypass) |
|------------------------|--------------:|-----------------:|
| PSNR                   |  **28.75 dB** |  20.21 dB        |
| SSIM                   |     **0.946** |     0.624        |
| SSIM (low-frequency)   |     **0.944** |     0.812        |
| SSIM (high-frequency)  |     **0.987** |     0.641        |
| MAE (lower is better)  |      **5.33** |    12.18         |
| Output resolution      |   2752×1536   |    1501×835      |
| SynthID verdict        |   not found   |    not found     |

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

| metric                 | gaussian (default) | edge mode |
|------------------------|--------------:|--------------:|
| PSNR                   |  **32.47 dB** |   31.47 dB    |
| SSIM                   |     **0.956** |     0.948     |
| SSIM (low-frequency)   |     **0.959** |     0.955     |
| SSIM (high-frequency)  |     **0.991** |     0.984     |
| MAE                    |      **3.82** |     4.08      |
| SynthID verdict        |   not found   |   not found   |

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

## How it works

```mermaid
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:

| file                                                      | size   | source |
|-----------------------------------------------------------|--------|--------|
| `qwen-image-2512-Q4_K_M.gguf`                             | ~13 GB | [Frederic75/Qwen-Image-2512-GGUF](https://huggingface.co/Frederic75/Qwen-Image-2512-GGUF) |
| `Qwen-Image-2512-Lightning-4steps-V1.0-fp32.safetensors`  | ~1.6 GB | [lightx2v/Qwen-Image-2512-Lightning](https://huggingface.co/lightx2v/Qwen-Image-2512-Lightning) |

### Run

```powershell
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

| flag                  | default | when to use |
|-----------------------|---------|-------------|
| `--seed N`            | random  | reproducible runs |
| `--denoise X [X X]`   | 0.25    | sweep denoise |
| `--steps N`           | 8       | per-pass step count |
| `--passes N`          | 1       | iterate img2img |
| `--restore-sigma X`   | 1.95    | tune detail restore |
| `--restore-mode M`    | gaussian | `edge` for shape-coherent contours |
| `--unsharp X`         | 0.0     | post-restore sharpen; 0.2 is the perceptual sweet spot |
| `--no-restore`        | off     | skip the frequency restore step |
| `--keep-intermediate` | off     | save the pre-restore clean output |
| `--transformer PATH`  | Q4_K_M  | try a different GGUF quant |

### Quality check

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

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

## Files

| file                                                      | role                                                       |
|-----------------------------------------------------------|------------------------------------------------------------|
| `desynth.py`                                              | main pipeline: img2img + restore in one call               |
| `compare.py`                                              | similarity metrics between two images                      |
| `embeds_cache.pt`                                         | cached prompt embeddings (~430 KB)            |
| `qwen-image-2512-Q4_K_M.gguf`                             | Qwen-Image transformer, GGUF Q4 quant (~13 GB)             |
| `Qwen-Image-2512-Lightning-4steps-V1.0-fp32.safetensors`  | 4-step Lightning distillation LoRA (~1.6 GB)               |

## 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](https://github.com/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.

DeSynth

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README