SynthID-Bypass¶

Development line: project:synthid-bypass · thread synthid-bypass
Last event: 2026-03-13 · 1 dated since 2026-03-13 · Researched: 2026-09-06 · confidence: medium

What it is¶

SynthID-Bypass is a community ComfyUI workflow for authorized robustness tests of Google’s image SynthID watermark, not a Google product.

• Current V2 runs as one graph: Qwen Image GGUF handles global reconstruction, and Z-Image Turbo handles face-specific cleanup.
• It is an image pipeline; Google documents separate SynthID modalities for text, audio, and video.

Development line¶

• 2026-03-13 — SynthID-Bypass v2.0 workflow artifact was linked. SynthID-Bypass and its v2.0 JSON were linked. No dated primary release note or immutable revision establishes a further project change on that day.

What changed¶

2026-03-08: A V2 announcement described one ComfyUI graph replacing V1’s separate workflow branches. It added resolution-aware denoise, a bundled custom-node pack, and a face-reconstruction path. 2026-03-13: SynthID-Bypass and its v2.0 JSON were linked. No dated primary release note or immutable revision establishes a further project change on that day.

How to use this¶

1. Define an authorized image-watermark robustness experiment. Retain source assets, outputs, and provenance together, because the project disclaimer forbids copyright circumvention and misrepresenting origin. — https://github.com/00quebec/Synthid-Bypass
2. For a disposable local evaluation setup, prepare a current ComfyUI installation, resolve listed V2 dependencies, and load the graph. Treat it as a workflow with external nodes and models, not a packaged application. — https://github.com/00quebec/Synthid-Bypass
3. Read a SynthID result as a likelihood signal: detected, not detected, or possibly detected. Preserve complementary provenance metadata, because non-detection does not prove human origin. — https://deepmind.google/blog/identifying-ai-generated-images-with-synthid/

Best practices¶

• Use only media we own or are explicitly authorized to test, and keep disclosure and provenance intact. The project disclaimer rejects copyright circumvention and misrepresentation of origin. — https://github.com/00quebec/Synthid-Bypass
• Treat V1 as history rather than current guidance. The repository marks the V1 folder as archived and keeps the active release at the root. — https://github.com/cebeuq/Synthid-Bypass/tree/main/V1
• Separate visual quality assessment from watermark detection. Google describes SynthID as a likelihood-based signal that complements metadata, and notes limits under extreme image manipulation. — https://deepmind.google/blog/identifying-ai-generated-images-with-synthid/

Superseded by this¶

• 2026-03-08: The V1 split into separate general, portrait, and GGUF workflows is legacy guidance. V2 was announced as the main single workflow and V1 as an archive.

Still unknown¶

• No immutable Git revision, release tag, or dated primary note was found for 2026-03-13. Current main-branch contents cannot prove the exact state on that day.
• The supplied https://hf.ru/linkd4c82 short link could not be safely resolved, so it is not evidence here.
• No independent test discloses detector version, input source, resolution, before/after artifacts, and reproducible settings. Available repository and secondary examples do not establish a bypass rate.
• The Simplified-Chinese source produced a secondary synopsis with only a relative publication label. No first-party Chinese material or independently reproducible Chinese test was found.

Sources¶

Agent brief¶

• Subject: project:synthid-bypass, thread synthid-bypass, 1 dated events 2026-03-13 → 2026-03-13.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.