SynthID in 2025: Where Google’s Invisible Watermark Shows Up (and Where It Doesn’t)
SynthID has become the clearest example of Google DeepMind’s approach to AI provenance: embed a machine-detectable signal directly into pixels (and video frames) so Google tools can later verify whether an image came from specific Google models. This article summarizes what SynthID is, which image and video generators actually ship it today, and how to reason about detection limits. All claims reflect public documentation as of November 23, 2025.
What SynthID is (in 90 seconds)
- Core idea: an invisible watermark added at generation time that Google tools can later read, with negligible visual impact.
- Scope: officially supported for Google’s own image and video models (Imagen family, Gemini image models, Veo) and for partner products that explicitly pipe to those models.
- Not a generic AI detector: if a model never wrote the SynthID signal, the verifier simply reports “not watermarked,” not “authentic.”
Fast snapshot: who actually uses SynthID
The only widely documented SynthID deployments are Google’s own image/video models and products that call them.
| Provider | Model / channel | SynthID? | Why it matters |
|---|---|---|---|
| Imagen 4 (incl. Ultra) via Vertex AI & Gemini API | Yes | Default, non-visible SynthID watermark on every generated image; verification API available. | |
| Gemini 2.5 Flash Image (“Nano Banana”) | Yes | All generations and edits carry SynthID for downstream verification. | |
| Gemini 3 Pro Image (“Nano Banana Pro”) | Yes | Same SynthID watermarking as Nano Banana, but higher-quality model. | |
| Veo (video) | Yes | SynthID stamped into every frame for video provenance. |
Who does not use SynthID (and what they use instead)
Most other major generators rely on C2PA metadata, visible labels, or proprietary schemes—not SynthID. Treat any “SynthID detected” message on these models as a red flag for your detection pipeline.
| Provider | Model / family | SynthID? | Alternative provenance |
|---|---|---|---|
| OpenAI | DALL·E 3; GPT-4o image; gpt-image-1 | No | C2PA metadata plus OpenAI’s own invisible signatures. |
| Microsoft | MAI-Image-1; Bing Image Creator | No | C2PA plus a visible corner watermark. |
| Midjourney | v5/v6 family | No | IPTC digital-source-type metadata; no SynthID. |
| Stability AI | Stable Diffusion 1.x/2.x/SDXL/3.x | No | Optional invisible_watermark or Stable Signature; not SynthID. |
| Black Forest Labs | FLUX.1 family | No | No public invisible watermark; metadata only. |
Partner and aggregator channels
SynthID appears in third-party products only when they explicitly route to Google’s models. The same UI can produce watermarked and non-watermarked outputs depending on the backend.
| Platform | Backend model path | SynthID? | Practical note |
|---|---|---|---|
| Adobe Firefly / Express | Gemini 2.5 Flash Image (Google mode) | Yes | Firefly also adds Content Credentials; SynthID comes from Google. |
| Adobe Firefly / Express | Firefly native models | No | C2PA-only; no SynthID signal. |
| NightCafe / Canva / Figma | Imagen 3/4 or Gemini image modes | Yes | When you pick the Google-backed mode, SynthID is embedded. |
| NightCafe / Canva / Figma | DALL·E, Flux, SDXL, etc. | No | No SynthID; provenance depends on each non-Google model. |
How SynthID detection works (and its limits)
- Signal location: encoded into pixel values (or video frames) rather than metadata, so it survives basic editing and format changes better than EXIF/IPTC tags.
- Verification: Google provides verification endpoints in Vertex AI and in the Gemini app; both look specifically for the SynthID pattern, not for general AI artifacts.
- False negatives: heavy transformations (extreme compression, aggressive filtering, image-to-image pipelines) can weaken or erase the signal.
- False positives: detection tools are designed to avoid them, but they also cannot confirm authenticity for non-Google models—they only say whether the SynthID mark is present.
Practical checklist for teams
- Map your generators: list which channels call Google models versus OpenAI, Midjourney, Flux, SDXL, etc. Only the Google branches will emit SynthID.
- Keep dual provenance: store both SynthID verification results and any available metadata (C2PA, IPTC) to cover mixed-model workflows.
- Be explicit in UX: when offering multiple backends, label which ones emit SynthID so downstream reviewers know what detection signals to expect.
- Test your edits: run typical resizing, cropping, and format conversions through the SynthID verifier to understand your internal false-negative envelope.
Metadata scrubbing is separate from SynthID. You can remove EXIF/IPTC tags with the Media Metadata Scrubber, but that does not erase SynthID signals embedded in pixels.
Key takeaways
- SynthID is a Google-specific invisible watermark. As of November 23, 2025, only Google’s own image/video models and partner integrations that call them embed it.
- OpenAI, Microsoft, Midjourney, Stability, Flux, Leonardo, Ideogram, Seedream, Janus Pro, Qwen-Image, CogView3, and similar models do not use SynthID; they rely on C2PA, visible labels, or custom schemes.
- Detection is binary: either the SynthID signal exists or it doesn’t. “Not watermarked” does not mean “human-made”; it simply means “not a Google SynthID output.”
Complete SynthID compatibility table (November 23, 2025)
For quick lookup, here’s the full Google DeepMind SynthID watermark matrix across major image/video models and partner channels.
| Provider | Model / family (image-focused) | Uses SynthID? | Evidence / notes |
|---|---|---|---|
| Imagen (2 / 3) on Vertex AI | Yes | Vertex AI docs: Imagen’s image generation and editing include a non-visible SynthID watermark by default and provide a verification API. (Google Cloud Documentation) | |
| Imagen 4 (incl. Imagen 4 Ultra) | Yes | Google: “All images generated by Imagen 4 models will continue to include a non-visible digital SynthID watermark.” (Google Developers Blog) | |
| Imagen family via Gemini API / Google AI Studio | Yes | Imagen API guide: “All generated images include a SynthID watermark.” (Google AI for Developers) | |
| Gemini 2.5 Flash Image (“Nano Banana”) | Yes | Google and third-party docs: every image created or edited with Gemini 2.5 Flash Image embeds an invisible SynthID watermark. (Google Developers Blog) | |
| Gemini 3 Pro Image (“Nano Banana Pro”) | Yes | Launch coverage and Google blog note that Nano Banana Pro outputs continue to carry SynthID watermarks. (Ars Technica) | |
| Gemini app: built-in image generator / editor | Yes (for Google-made images) | New Gemini app feature can detect whether an image was generated/edited by Google AI by scanning for SynthID; it only works reliably on Google-model outputs. (The Verge) | |
| Google Photos – Magic Editor & other generative photo tools | Yes (for edits) | Google adds SynthID watermarks to images edited with Magic Editor’s generative features in Photos to flag AI-manipulated pictures. (The Verge) | |
| Veo (video generation) | Yes (video frames) | Veo and Imagen 3 on Vertex AI are documented as embedding SynthID into every image and video frame. (Google Cloud) | |
| Older Imagen variants on Vertex AI (Imagen v1, early preview) | Yes (where watermarking supported) | Early SynthID launch: SynthID initially rolled out specifically for Imagen on Vertex AI to watermark and later detect its outputs. (Google DeepMind) | |
| Adobe (Google-backed mode) | Gemini 2.5 Flash Image / Nano Banana used inside Adobe Firefly & Adobe Express | Yes | Adobe design blog: when you select Google’s Gemini 2.5 Flash Image inside Firefly/Express, outputs are invisibly watermarked via SynthID in addition to Adobe’s Content Credentials. (AI Graphic Design Courses) |
| Adobe (native models) | Firefly Image models (Firefly 1/2 etc., incl. Photoshop Generative Fill/Expand) | No (SynthID); uses C2PA/Content Credentials | Adobe’s provenance stack relies on Content Credentials (C2PA) metadata, not SynthID; SynthID appears only when they pipe to Google’s models. (The Verge) |
| OpenAI | DALL·E 3 | No SynthID; uses C2PA + sometimes invisible OpenAI watermarking | OpenAI and coverage state DALL·E 3 images carry C2PA content credentials and OpenAI’s own invisible signatures; nothing indicates SynthID usage. (OpenAI Help Center) |
| OpenAI | GPT-4o native image generation | No SynthID; uses C2PA + OpenAI marks | OpenAI’s 4o image blog: all GPT-4o images ship with C2PA metadata identifying the model; later reporting discusses OpenAI testing visible and invisible watermarks, again not SynthID. (OpenAI) |
| OpenAI | gpt-image-1 (API image model) | No SynthID; C2PA only | API docs: gpt-image-1 reuses 4o’s safeguards including C2PA metadata; no mention of SynthID, and industry comparisons explicitly contrast OpenAI’s C2PA approach with Google’s SynthID. (OpenAI) |
| Microsoft | MAI-Image-1 (first-party image model) | No SynthID; visible mark + C2PA | Bing Image Creator and MAI-Image-1 coverage: Microsoft uses a visible corner watermark plus C2PA content credentials for provenance, not Google’s SynthID. (Microsoft) |
| Microsoft | Bing Image Creator (MAI-Image-1 / DALL·E 3 / GPT-4o backends) | No SynthID | Same as above: Bing’s provenance relies on built-in watermark + C2PA metadata regardless of backend model. (Microsoft) |
| Midjourney | Midjourney v5/v6 family | No SynthID | Google and independent SynthID explainers: SynthID detectors simply report “not watermarked” for images generated by Midjourney; it uses IPTC “digital source type” metadata and/or its own tagging, not SynthID. (DEV Community) |
| Stability AI | Stable Diffusion 1.x / 2.x / SDXL / Stable Diffusion 3.x | No SynthID | SynthID documentation and third-party testing: images from Stable Diffusion are treated as “not watermarked” by SynthID; any watermarks are either optional libraries (e.g. invisible_watermark) or distinct schemes like Meta’s Stable Signature, not SynthID. (DEV Community) |
| Black Forest Labs | FLUX.1 family (flux-schnell, flux-dev, Flux Kontext, etc.) | No SynthID | Guides comparing Flux vs Nano Banana explicitly note that Flux and other non-Google generators do not embed hidden watermarks; SynthID is specifically tied to Google’s stack. (CyberSEO Pro) |
| Leonardo.ai | Leonardo diffusion models | No SynthID (no public evidence) | Leonardo’s docs and community threads discuss training data watermarks and occasional visible marks in generated images, but there is no mention of SynthID or Google integration. (Reddit) |
| Ideogram | Ideogram (v2–v4) text-to-image models | No SynthID | Ideogram materials mention optional visible watermarks on free plans and generic metadata tagging; no linkage to SynthID, and SynthID detectors treat non-Google images as unwatermarked. (Digital Software Labs) |
| Adobe (other) | Photoshop / Illustrator Generative tools using Firefly only | No SynthID; C2PA only | Adobe’s Content Credentials initiative is based on C2PA; SynthID only appears when a Google model is explicitly selected as a partner model. (The Verge) |
| ByteDance | Seedream 4.0 | No SynthID (no public spec) | Comparisons with Nano Banana note that Seedream 4.0 has no publicly documented model-specific watermarking like SynthID or C2PA; reviewers explicitly flag this as an uncertainty. (TechRadar) |
| DeepSeek | Janus Pro (Janus-Pro-7B, etc.) | No SynthID | Janus Pro is an independent Chinese image model; SynthID surveys list Janus Pro as a type of model SynthID might in theory detect if those vendors adopted it, but there is no indication DeepSeek has done so. (Reuters) |
| Alibaba / Qwen | Qwen-Image family | No SynthID (no public indication) | Qwen-Image technical report and API docs focus on architecture and capabilities; there is no mention of SynthID or DeepMind integration. (Qianwen) |
| Zhipu AI | CogView3 / CogView-3Plus | No SynthID | CogView3 is open-source; papers and community implementations mention standard or optional invisible-watermark libraries, not SynthID. (ComfyUI Wiki) |
| Aggregators / multi-model platforms | NightCafe, Canva, Figma, etc. when they call Google Imagen/Gemini | Yes for the Google-backed modes only | When these platforms use Imagen 3/4 or Gemini image models, the underlying Google model applies SynthID; when they use DALL·E, Flux, etc., those images have no SynthID and instead rely on each model’s own provenance scheme. (Zapier) |
| Aggregators / multi-model platforms | Same tools when they call OpenAI, Flux, SDXL, etc. | No SynthID | SynthID detection docs and ecosystem articles repeatedly stress that non-Google model outputs—including from OpenAI, Microsoft, Midjourney, Stability, Flux, etc.—do not contain SynthID signatures. (Chrome Unboxed) |
Sources: Google DeepMind and Google Cloud SynthID documentation, Google Developers Blog (Imagen 4; Gemini 2.5 Flash Image), Google AI for Developers (Imagen API), Ars Technica coverage of Gemini 3 Pro Image, Verge reporting on Gemini app and Magic Editor, Android Central on mobile detection, and cross-vendor provenance comparisons.