Multi-modal watermark dataset contribution: 1,113 files across image/audio/video (SynthID + C2PA)

[biangacila]

Multi-Modal Watermark Dataset: 1,633 Files

Download: dataset.tar.gz (1.3 GB)

---

Summary

Modality	Model	Watermark	Reference	Diverse	Total
Image	gemini/nano-banana-pro-preview	SynthID	390	100	490
Image	gemini/gemini-3.1-flash-image-preview	SynthID	489	100	589
Image	openai/dall-e-3	C2PA metadata	180	100	280
Audio	gemini/gemini-2.5-flash-preview-tts	SynthID	187	30	217
Video	gemini/veo-3.1-generate-preview	SynthID	18	10	28
Video	gemini/veo-3.0-fast-generate-001	SynthID	6	—	6
Video	openai/sora-2	C2PA metadata	13	10	23
			1,283	350	1,633

Only models with confirmed watermarking were used. Models without confirmed watermarks (OpenAI TTS, gpt-image-1/1.5, FLUX, DashScope) were intentionally excluded.

---

Image — 1,359 files

Nano Banana Pro SynthID (490 files)

The model the project specifically requests for contributions. Generated via image-to-image (send pure color PNG → NBP recreates with SynthID).

Reference (390): 6 colors × 7 resolutions × ~10 each.

Color	RGB	Purpose
Black	#000000	Watermark ≈ entire pixel content
White	#FFFFFF	Cross-validation via phase inversion
Red	#FF0000	R-channel carrier isolation
Green	#00FF00	G-channel carrier isolation (strongest per project findings)
Blue	#0000FF	B-channel carrier isolation
Gray	#808080	Mid-tone carrier behavior

Diverse (100): Text-to-image with varied prompts.

Gemini 3.1 Flash SynthID (589 files)

Same approach as NBP but from a different Gemini model — allows comparing SynthID patterns across model variants.

Reference (489): 6 colors × 7 resolutions × ~10 each.

7 distinct output resolutions (each has different SynthID carrier frequency positions):

Output Resolution	Input Size	Count (NBP)	Count (Flash)
1024×1024	1024×1024	53	67
1195×896	640×480	49	71
1264×843	768×512	56	72
1365×768	1920×1080	53	69
1440×720	2048×1024	56	63
720×1440	1024×2048	54	72
768×1365	1080×1920	53	71

Diverse (100): Text-to-image with varied prompts. Primarily 1408×768.

DALL-E 3 C2PA Baseline (280 files)

Control group — C2PA metadata watermark (trivially removable via exiftool -all= image.png).

Reference (180): 6 colors × 3 sizes (1024×1024, 1024×1792, 1792×1024) × 10 each.
Diverse (100): Mixed sizes.

---

Audio — 217 files

Gemini TTS SynthID

The only confirmed audio watermarker. Same principle as black images: minimal utterances maximize watermark-to-content ratio.

Reference (187): 8 voices × 10 texts × 3 repeated generations (for differential extraction).

• Voices: Kore, Puck, Charon, Fenrir, Aoede, Leda, Orus, Zephyr
• Texts: "A.", "One.", "Ah.", "Hmm.", "Okay.", "Yes.", "No.", "Hello.", "The.", "Aaaaaaaah."
• Format: WAV, 24 kHz, 16-bit mono, 0.5–2.9s

Diverse (30): Natural sentences, 4.5–7.6s.

---

Video — 57 files

Veo 3.1 SynthID (28 files)

Reference (18): Solid color screens × 3 each, 4s, 1280×720.
Diverse (10): Natural scenes, 8s.

Veo 3.0 Fast SynthID (6 files)

Reference (6): Black/white, 4s.

Sora 2 C2PA Baseline (23 files)

Reference (13): Solid colors (some blocked by moderation). Diverse (10): 8s.

---

Directory Structure

dataset/
├── image/
│   ├── nano-banana-pro/
│   │   ├── reference/    # 390 PNGs — {color}_{WxH}_{idx}.png
│   │   └── diverse/      # 100 PNGs — diverse_{idx}.png
│   ├── gemini/
│   │   ├── reference/    # 489 PNGs — {color}_{WxH}_{idx}.png
│   │   └── diverse/      # 100 PNGs — diverse_{idx}.png
│   └── dall-e-3/
│       ├── reference/    # 180 PNGs — {color}_{size}_{idx}.png
│       └── diverse/      # 100 PNGs — diverse_{size}_{idx}.png
├── audio/
│   └── gemini-tts/
│       ├── reference/    # 187 WAVs — ref{NN}_{voice}_{idx}.wav
│       └── diverse/      #  30 WAVs — diverse_{voice}_{idx}.wav
└── video/
    ├── veo-3.1/
    │   ├── reference/    #  18 MP4s — {color}_{idx}.mp4
    │   └── diverse/      #  10 MP4s — diverse_{idx}.mp4
    ├── veo-3.0-fast/
    │   └── reference/    #   6 MP4s
    └── sora-2/
        ├── reference/    #  13 MP4s
        └── diverse/      #  10 MP4s

Watermark Evidence

Model	System	Evidence
Nano Banana Pro / Gemini Flash / Veo / Gemini TTS	SynthID (invisible, frequency-domain)	deepmind.google/technologies/synthid
DALL-E 3 / Sora 2	C2PA metadata (XMP, easily stripped)	c2pa.org

Value

1. Nano Banana Pro data — the specific model the project requests for contributions
2. Two Gemini model variants (NBP + Flash) — compare SynthID patterns across model versions
3. 7 new resolution profiles — existing dataset only has 1024×1024 and 2816×1536
4. First audio watermark dataset — enables SynthID audio detection/removal research
5. First video watermark dataset — enables SynthID multi-frame spectral analysis
6. Per-channel R/G/B reference — direct per-channel carrier isolation
7. C2PA baseline — control group for invisible vs. metadata watermarking
8. Differential audio pairs — same text × voice × 3 gens for cross-correlation extraction

[aloshdenny]

Hi

Thankyou for contributing this dataset!
I have created a huggingface dataset repo where you can raise PRs in a structured manner: https://huggingface.co/datasets/aoxo/reverse-synthid
I noticed that the images are of varying dimensions so I'd appreciate if you could group images in the order of: Provider (Gemini/Dall-E/etc..) > Resolution (1024x1024, etc..)

[BensonRen]

Great contributions! wondering did any of your tried the detection methods for the diverse set there? I have faced some trouble with the robustdetection method as it does not detect the real distribution of diverse generated image well.

[biangacila]

HuggingFace PR: https://huggingface.co/datasets/aoxo/reverse-synthid/discussions/6